Version History

Every trained voice kept talking after its words because training paired each recording with a sentence from and older script that was never shown.

The previous save of this name carried no code: the change never applied and the check did not notice.

The third training run of the day was ended by the same background reload of the language model, because the check that asks whether a run is under way can time out under load and was read as no.

A trained voice sometimes kept making sound after a line was said.

The Test Voice button on the Voice card failed with a script error as soon as the sliders were added. It plays the test line again, with the sliders as they stand.

The voice now served is today's trained one and it is remembered across restarts.

The Voice card in the admin panel has three new sliders: how fast he speaks, how high or low, and how much his delivery varies.

Starting a voice training run now holds the whole graphics card for the run.

Leaving the voice training wizard part way through no longer sends you back to the first sentence.

When a reply is cut short, he now remembers the part that was actually heard instead of forgetting the whole reply.

Setup now asks whether the person is setting this up for themself or someone else is doing it for the family.

When someone asks more than once whether he is real or really the person, each answer is now in different words, and from the second time he asks what is making them wonder.

After the passing is recorded, the first time each person talks to him the screen shows what he is and is not, in the persona's own words when they were written, and waits for a tap; a quiet line stays on the first screen.

He now uses a nickname at most once in any three replies in a row, across greetings, goodbyes and maths or coding help, instead of in nearly every reply as before.

When a reply states something the notes do not hold, he is asked to answer again from what is written.

Asked whether they have something nothing is written down about, such as a young child asking 'do I have a teddy bear?', he now says he does not know of one and asks them about it, instead of answering about himself or offering something else in its place.

The live check of how changes in the family's life are tracked now removes everything it writes, including rows the ordinary screens protect from deletion, and reports any row it failed to remove instead of declaring the store unchanged.

The check that a relationship ending is remembered as and ending now runs as a live scenario against the real store: a child's boyfriend is recorded, the child says it ended, the conversation closes, and the parent asking afterwards hears that it is over.

A test conversation in which a child shares news about a first boyfriend, asks for discretion, and later says it has ended found six things to put right.

When he explains a program that prints your name, or repeats what someone wrote on a card, the words inside the quotation marks and inside code now stay exactly as written.

An ordinary conversation was walked through the last month of changes and three things were put right.

Saving a change on whichever project is not currently running no longer stops halfway: the check that looks for the running system is allowed to find nothing.

A saved snapshot now includes the memory store only when this device's own store is the one running, so another household's memories can never end up in it.

The weekly backup now archives the persona's configuration, birthdates and questionnaire answers, only ever exports its own memory store, and catches up on its own when the device was off at the scheduled time.

A review of the last lesson mode fixes closed six gaps.

Four more of the lesson mode fixes, each read twice by a second reader who tried to break it.

Five of the twelve fixes from the lesson mode test, each checked by a second reader who tried to break it.

The maths, coding and game modes were put through eighteen full conversations: how each starts, whether it stays with the subject through several follow ups, whether it follows a change of topic or language inside the same session, what happens when the child drifts off topic, and how it ends.

When someone tells him about their day he listens instead of reaching for a memory, which is right.

When someone asked for a memory of the persona and one particular family member, the question was treated as a request for any memory at all and answered from a handful of the asker's own, picked at random.

A nine year old said hello and was asked whether she remembered babbling as a baby.

The advice that makes the difference between a natural voice and a synthetic one was only ever said out loud, never written where someone recording would see it.

The trained voice sounded like the right person and still sounded synthetic, and the recordings explain why: a hundred and forty clips, eight minutes in total, most of them three or four seconds, and only two longer than six.

Deciding which memories to bring into a reply was one set of instructions trying to serve two opposite situations, and it could not.

A two year old said his shoes were too small and was told about his sister as a baby.

After training finished, pressing Test produced no sound.

Finishing voice training used to change nothing at all — the newly learned voice was written to disk and then never spoken, while the system reported success.

If the system restarted while the voice was being trained, the training itself carried on but nothing could stop it any more, and the graphics card stayed occupied with no way to get it back.

Voice training would not start.

Opening the demo site asked for its password over and over, and no amount of typing it would settle.

In the personality card, the sentence describing what each slider actually does was being shown in a completely different typeface and about twice the size of everything around it, pushing the rows out of shape.

On the demo site, opening the admin panel asked for the site password, accepted it, then asked again in a loop, and the PIN screen never appeared.

One sentence out of a hundred and forty would not save, no matter how many times it was read, and the screen gave no reason.

The Voice Training card in the admin panel was drawing and empty box — no title, no button, no error.

The memories a full memory reset would spare — questionnaire answers, seeded memories, and anything written by hand in the memory editor — can no longer be deleted or edited from the ordinary screens, and a bulk delete skips them however the filter is set, saying how many it kept.

Where a memory came from used to be a dropdown, so it could only ever answer 'show me this one kind'.

A memory is filed under the name of the person it is about, and the system was reading that as the person who said it.

Housekeeping only: the two projects share one version number and had drifted apart by one when a save was interrupted. No change to how anything behaves.

A memory containing a less than sign a note about a score, and arrow, anything with a bracket used to lose the words after it in the memory timeline, because the display read them as formatting instead of text.

When a conversation ended, the persona was filing the questions someone asked as though they were facts about that person's life three of the ten things learned from the children's conversations were entries like 'asked about photos and videos' or 'asked about when he moved'.

A child told her father that a friend at school had been unkind to her.

From now on, fixing how the persona speaks, listens or decides requires adding the check that would have caught the fault, in the same save.

Two things that were still wrong are fixed.

A tool now reports how often the tidying that happens after the persona writes a reply actually changes anything, and which rule made each change.

About one reply in five is rewritten after the persona has written it, and half of those rewrites were damage.

The check that notices when the persona states something his notes do not contain has existed for a long time, but it only ever ran after the reply had been delivered, and only when a rough first pass had already spotted a date or a name in capitals.

Three repairs to how people are named.

Two leftover backup files were still being carried in the project.

The Person Settings screen still showed the old nought to a hundred sliders after the other two screens had moved to 1 to 5.

A personality setting changed from outside the running system looked like it had been saved and was quietly undone a few minutes later, twice.

Each personality trait has exactly five settings, because five different instructions exist for it and no more.

Answering the child about her friend, the persona kept reaching for one line that he wished he could be there to help in twelve replies out of twelve, always in the same position, while everything around it changed.

A child told her father that her best friend at school is mean to her, and he answered with the same hollow line twice: that he wished he could be there to help.

A child asked whether the hospital she was born in was the one her father was born in, and he agreed naming her hospital as his own, five times out of five.

Two changes to how replies reach the family.

Replies about a memory are written by the persona from the note again, in his own words.

The date written on a memory is now spoken only on the turn that actually asks for a time when did that happen, what year, how long ago.

The Backend Offline badge on the main screen is now a button.

Asked when he came to the country, he answered with the city and never the date — even though the date is written down and he had it in front of him.

Asked when he came to the USA, he answered with a place — the city, not the date — and gave the right date immediately when the same thing was asked as 'what date was that?'.

Asked what he did for fun, he answered the youngest with a road trip, a jeep trail and the shooting range in one sentence, then explained to a two year old what a gun range is.

Reverted a change from last night that made him lose the thread.

Three ways the persona's own words could come back as evidence are closed.

Two follow ups to the last change, both measured over several runs because a single pass proves nothing.

Five turns with the youngest child showed four separate ways a correct answer could be lost between the memory and the reply, and all four are fixed.

The thinking process map no longer shows children the step where something they said is passed to a parent.

The little map in the header now draws properly when the header is opened, and tapping it opens the same movable window the debug panel uses, over the conversation and out of the way, instead of a second full screen panel.

The thinking process map now names what is happening rather than who: reading the message, picking the words, meaning search, the memories, sorting by fit, putting it together, thinking, checking for invention, tidying the words, speaking.

The thinking process map is smaller and out of the way, each part explains itself when you touch it rather than being listed underneath, and its colour now tells you how things are going: the usual cyan while answers arrive at their normal speed, amber when a turn is slower than usual for this machine, orange when it is taking a long time, red if something failed.

The map in the header is now called the thinking process, with a line under it saying what it is: what happens between your words and the answer.

A small live map of the persona's thinking now sits at the end of the status row when the header is open, lighting up as it reads your message, searches its memories, decides which ones answer you, writes the reply and checks it.

The synapse map now fires on every call the system makes, including the repeated ones the written trace does not print, such as each memory set aside as not answering the question.

The reasoning window keeps every decision but no longer repeats itself: a line that carries nothing readable is left out, and a step that runs once per item, such as setting aside each memory that did not answer the question, is shown twice with the total rather than fourteen times.

The reasoning window and the synapse map now show every decision the system makes on a turn, not the fraction that happened to be wired up: the guides that fire, the guards that send a reply back to be rewritten, the notes added to the prompt, the privacy filtering.

Requests about how the persona speaks are now understood the way a person would understand them, whatever the wording, instead of by a list of phrasings: the persona asks itself one question about what was meant, and only a real request to change its manner changes it.

When a family member talks about themselves or reports what someone else said, such as 'do you think I'm too blunt?' or 'the teacher says you're too harsh', the persona treats it as conversation rather than as a request to change how it speaks.

You can now adjust how the persona speaks by simply asking, the way you would ask a person: 'can you be a bit more sensitive?', 'you're being too harsh', 'keep it short', 'be a bit warmer', 'less jokes please'.

The persona no longer tells you what you must be feeling.

The personality sliders now do what their ends say.

Two fixes found while measuring the personality sliders.

The persona's own relatives who exist only in its memories, such as a sister, and uncle or a parent, are now known to its honesty check as well: when it tells a child about 'your aunt' or 'your great uncle', or a spouse about 'your sister in law', those words are recognised as the people its memories name instead of being treated as invented.

Saving is safer when both installs live on one machine: if the check for the other install's running containers fails for any reason, it is now treated as running, so a second stack is never started beside the first.

The persona now knows who is who to whom for the whole household, not just for a spouse and children.

Better answers to a child's questions about memories.

Pictures now appear only when someone asks for them by name or when the persona is actually showing them.

Fixed the album opening by itself when a child asked a question right after pictures had been mentioned.

Improvements found by testing the customer instance, which has far fewer memories than the family one.

Two things learned from one real conversation.

A second measured pass over response speed, with every result on the record.

A measured pass over what the same graphics card and model still had in them, with every result on the record.

The engineering record notes that two face folders from this household had been carried over to the twin install when it was created, and that they have been removed there.

The engineering record now carries the outcome of the twin install's first full test run and the family's own re run afterwards, with every suite passing here.

Two rules were tightened: a child speaking alone is again asked about a parent's serious illness she reported, and a young person on the road back is again offered the day count; those only wait when a child is in the room beside the speaker.

A look now does what the request asks.

The parts of the system that decide what a child may hear, which memories stay private, when a safety gate steps in and when a parent is alerted were writing their notes to a logger nobody had connected, so none of it ever appeared in the log or in a support bundle.

When two known people are in front of the camera, the conversation now stays with the person who has it while they are still there.

When a child showed the persona something and said "let me show you my drawing", the memory was kept word for word as "showed me my drawing", which read back as the persona's own drawing.

When a family member says let me introduce you to my boyfriend, my friend or my teacher, the persona looks the same way it looks at a drawing, describes the person plainly without ever guessing a name or and age, says hello, and asks who they are.

When a child asked where the keys were, the persona treated it as a general knowledge question and answered without looking at what the family had told him about the keys, so he made something up about them hiding.

When a family member says something like let me show you my drawing, the persona now asks them to hold it up to the camera and say when.

Housekeeping after the model change.

The small model that decides what a message is, files what is worth remembering and answers the help panel has been replaced by one that can also look at pictures, chosen by measurement against the old one and against the model that described photos until now.

In the hardest conversations the persona had started using the spouse's given name, which is not how he speaks to her.

A live transcript of the hardest days found four things the checks had not, and all four are fixed.

The two answers in the questionnaire section When it is hard are now used as a guide for how the persona formulates what he says in those moments, never as lines to repeat, and they are kept apart from his memories so they cannot come up in ordinary conversation.

The life questionnaire in the admin panel has a new section, When it is hard, with two questions: what he would say to someone he loves who has been given news that may not end well, and what he would say to a grown child in rehab, after a relapse, or counting sober days.

The persona now follows a family member through drink or drug trouble the way a loving parent would.

The persona now understands the gravest kind of health news.

The persona now understands how serious a health issue is and follows up on it.

The system now records exactly which AI models it runs on, by their fingerprints rather than their names, together with the versions of the software around them.

The assistant now learns from conversation what each family member likes and dislikes, what they are into, and how they want to be spoken to, such as 'please don't ask me about school'.

Follows on from the previous release.

The assistant now keeps track of things that change.

Housekeeping only.

Repairs a fault introduced by the previous update that stopped the assistant from starting until it was patched by hand.

Tried giving each language model two request slots so background work could overlap a reply instead of queueing behind it.

The system was holding three language models in memory at once.

A child asking her father where he was born was sent to the other parent as and urgent alert, three times over.

A memory holding several videos showed only one of them, and that one wore a photograph of something else as its picture — the cruise memory has four videos, and three could not be reached at all.

The settings panel had a scrollbar down its side and, after the audio controls were pinned into that same scrolling area, a second one along the bottom that covered the last row of controls.

Two ways a child's own words came back wrong.

The list of sessions to pick up again had grown unhelpful: one conversation left several entries behind, titled with whatever was said partway through — 'it was pepperoni', 'that was a good one' — because each entry took its name from a moving window rather than from the question that started it.

Asked about the persona's own mother right after another question, the answer came back about the wrong woman — sometimes about the children's mother instead, once refusing outright with a line that made no sense.

Asked where their mother was born, the children were told someone else's birthplace — and aunt's, in one case, with her photograph attached — and in another a famous city that appears in no memory at all.

The questionnaire area now asks about the whole family, not only the persona.

Seventy written anniversary letters had no way to ever be delivered — they carried no trigger of any kind and could only fire if a conversation happened to mention the right words.

The memory store has always been mostly the persona's own life; the people it was built for barely exist in it.

Keeping every missed birthday indefinitely had one unintended edge: messages written for ages a child had already lived — authored retroactively so the collection covers every year — counted as missed and would all have arrived at once on the next hello, four baby birthday letters to a nine year old.

If more than one birthday message is waiting — say two years passed without a sign in — they all arrive together on the very next visit, oldest year first, each one telling honestly which birthday it has been keeping for, and each marked as delivered only once it has actually been said.

Birthday messages no longer expire at all — they wait until the person actually opens the chat, however long that takes, and a message that waited says so honestly, naming which birthday it has been keeping for.

Two holes in how milestone messages are delivered.

Three parts of the mind that were recently worked on had no circle on the live synapse map: the pointer that aims at the answering sentence inside a long memory, the correction loop that takes in a family member's corrections, and the consolidation that turns finished conversations into memories.

Two gaps in what children could be shown.

Asking how someone in the family is doing today — his sister, his father, the children's mother, a grandparent — was being answered with invented reports: a sister denied, a grandmother's morning of babysitting that never happened, the children's mother described as busy and stressed.

Asked how his mom was, he answered about the children's mother instead — inventing that she was busy with work and stressed, things nobody ever recorded.

When a memory written about both children was retold to one of them, the names were adapted but the small words around them were not — producing sentences like 'getting you and your sister down for their naps, I'd lie with them' — and the reply came out garbled trying to smooth that over.

A short correction naming a place — 'no, in cancun' after asking what you ate — was answered from nothing, because messages of three words or fewer skipped the memory search entirely on word count alone.

Three fixes for how test conversations are cleaned away.

The system has always had the machinery to learn from conversations — a nightly pass that summarises each finished session and keeps anything worth remembering.

Asked where his wife was born, he gave his brother's birthplace instead, and on another attempt said he did not know at all.

Photos still appear with every memory that has one — that part is unchanged.

Photos in the conversation were allowed to overlap into the messages below them, so every reply after a picture was a different width and the text edge moved around as more pictures appeared.

He was ending nearly every reply the same way — asking what your favourite part was, three turns running, the last time asking about something you had just answered, and announcing a photo in every single turn of and exchange, twice in identical words.

The tool that measures answer quality was marking correct answers wrong.

Tested whether sending fewer memories with each question would improve answers.

Added the large mixture of experts model to the comparison tool and measured it.

Added a tool that grades each language model already installed on the same set of questions with known answers, reporting how often the reply carries the fact, how long the wait is, and how much it writes.

The tools that measure how often a memory is used were writing questions that could not be answered correctly.

When a child asked where they were born, the answer came back as the persona's own birthplace, describing his birth as though it were theirs.

Added two measurement tools.

When and answer sat inside a long memory, the reply often picked the wrong detail from it.

When a family member's own words about a day the persona also lived through came back at the top of a search, his reply treated the whole day as theirs alone — asked about a holiday they took together, he answered as though he had only been told about it.

When someone said something about themselves — 'my head hurts', 'I got in trouble', 'the dog bit me' — it was being treated as a question to be answered out of the persona's own recorded life, so his own memories could be offered as the explanation for their day.

Questions about your shared past that did not happen to say 'we' or 'our' — such as asking what a hotel was called — were being treated as general knowledge questions.

An attempt to make answers less clipped was measured and rejected, and the reason is written into the code so it is not tried again blindly.

Asking what the hotel in Cancun was called got the name of a hotel he never went to, four times out of four, while the note saying where they actually stayed sat in front of him.

No change to how he answers.

Asking what he put on a classmate's seat at school was answered with and offer to play different music, because the words put on were read as a request to play something and the rest of the sentence became the name of a song.

The same question used to answer well one time and poorly the next, because every reply is composed fresh each time.

A child asked what happens when someone dies and was told that her brother's love was still with her even though he was no longer here.

The deep thinking switch has been taken out of the settings.

Four different places named the model that answers the family, and they disagreed: the code said one, the saved admin settings said another, the container was started with a third, and forty eight files read the container's value while falling back to the first.

Starting a fresh conversation now really starts fresh.

Saying the word for something the system can do no longer counts as asking for it.

Answers he typed into the life questionnaire are stored as he wrote them, with the question kept separately, so and answer that only makes sense beside its question could be read out as though it were a reply to whatever was just asked.

Seven notes about difficulties between the parents are now kept for the adults, so the children do not meet them while they are young.

Notes kept for the adults now open to the children when they are grown, at eighteen, instead of being sealed from them for good.

A private note about the parents was read out to a child who had asked about a family holiday.

A message that begins with a small word like oh, and, so or wait was being treated as and aside that needed no memory at all — so saying oh, i meant the disney cruise, when did we get there was answered with I don't have that one any more, about the memory named in the same sentence.

Asking whether he can see why you are upset now gets and answer, instead of the comfort written for someone who has just said they are upset.

Asked about the ice cream on the ship, he said he did not remember any — from a memory that reads we had ice cream whenever we wanted, about a week aboard one.

Photographs from other trips no longer appear alongside the one being talked about.

When he reached for the wrong term of endearment, the word was simply removed and the sentence ended up addressing nobody at all: saying he did not have something written down, to no one.

He no longer describes his own remembering in the language of a filing system.

A memory can be a paragraph long, and asking about one moment inside it did not always reach that moment: asked about the constellation on a bunk bed ceiling, he answered about a baby babbling, from a memory that says nothing about babbling.

The memories list can now be filtered by origin: what somebody wrote down on purpose, and what was kept automatically from a conversation.

When someone reminisces they repeat back what they are looking at, and that reply is kept as a memory of its own.

Asking about one moment from a memory that holds many photographs now opens on that photograph instead of the first one in the album.

The place field for a photograph now accepts coordinates as well as a name.

The media panel now shows where each photograph was taken and lets you correct it while you review.

A small correction to the picture addresses used by the media panel: each one already carries a version stamp from the file itself, and a second stamp was being appended with the wrong separator, producing and address with two question marks in it.

The picture viewer in the media panel now has Previous and Next buttons, with a counter showing where you are in the collection, so photographs can be reviewed and described one after another without closing and reopening the viewer each time.

Asking for a picture of someone by how the family actually refers to them now finds the right photograph.

Photographs no longer reload every time you look at them.

Memories that hold many photographs now show the one that fits what is being talked about, instead of always the first.

During a long, fond conversation about a trip, no picture ever appeared.

A short follow up now keeps the subject of the conversation instead of being searched on its own words.

The hardware panel no longer contradicts itself.

The warning triangle about the graphics card slowing down no longer appears when nothing is wrong.

The photos on a memory can now be dragged into place, with a finger or a mouse, instead of nudged one step at a time moving the last photo of a hundred to the front took a hundred clicks.

The memory editor can now arrange a memory's photos into any order with arrows on each one the first is the picture the chat shows and the gallery opens on instead of leaving them in whatever order they were picked.

The reasoning model the system runs has a thinking mode that was never being used: it is requested per message, and the system never asked for it, so the thinking model and the non thinking one behaved identically measured turn for turn on the same conversation.

Asked to tell the story of a day rather than asked a direct question, the persona was changing the numbers it had written down a child's birth weight came out three ounces light on one telling and two on the next, and once with the other child's figures entirely.

Re measured how often the persona states something its notes do not contain, using the check that was fixed yesterday the earlier figures were taken with a check that could not see invention at all.

When a child owns up to something taking a toy, breaking something the persona now guides them to put it right.

Asking the persona to tell you more about something now continues the same story instead of starting over or falling silent previously 'tell me more' searched for those three words, found nothing, and the persona answered that it did not remember the trip it had just described.

When a question names a holiday, the persona no longer describes a different day as that holiday asking about Easter returned a birthday party at a farm, described as Easter.

Asking about a holiday halloween, christmas, easter, thanksgiving returned nothing at all, because naming a holiday was being read as a date and the search was narrowed to a few days around last year's occurrence.

Swept all stored memories and rebuilt their keyword lists with the corrected rules, then confirmed a second pass changes nothing.

Memory keywords were being stored far too generously: one long memory carried 64 of them, most simply repeating words already in the memory's own text, and the list hit its limit so the end of that memory had no keywords at all.

When a child asks how to fix something, the persona can now actually walk them through it.

Four fixes from one conversation.

Three fixes from a real conversation with the youngest.

While the interface shows the thinking indicator, the language model node on the synapse map now visibly works glowing and receiving steady impulses from the prompt instead of sitting frozen.

The synapse map now fires along its output path for every word as it appears on screen, for the whole time the reply is being written out previously the pulses tracked network delivery, which finishes in a moment, so the map sat quiet while the family watched the words arrive.

The live synapse map now includes the chat input itself as a station: pressing send fires a visible signal from the person through the input into the pipeline, and each word of the streamed reply pulses down the output path to the screen as it is displayed.

The debug view gains a second window: a live synapse map of the entire reply pipeline.

Asking about a place now finds the memories that happened there, even when the memory's text never names the town the place stored with a photo or typed by hand now travels with the memory into search, into the relevance decision, and into the reply itself.

When someone suggests the persona is a program or asks who programmed it, the reply now firmly and warmly rejects the idea instead of cleverly playing along with it.

Telling the persona 'you are a program' or asking 'who programmed you' no longer switches the conversation into coding helper mode or returns a lost memory apology the persona now answers such questions as itself, staying in character.

Memories that came from the life questionnaire now display the question they answered, shown as context above the answer in the memory browser a short answer like a clothing description no longer appears without explanation.

Asking whether a picture or other belonging exists no longer comes back with music search results music claims the question only when it is really about music or names a song, artist, or playlist that actually exists in the library.

The instant child safety replies now fire even when the child's message has spelling mistakes, doubled letters, or words run together previously a single typo could slip a worrying message past the immediate response, leaving only the slower background check.

The conversation window on a parental alert now shows the persona's reply to the flagged message, in order and without cut off sentences, so a parent sees the whole exchange around the line that raised the alert.

The immediate safety reply a child receives after saying something alarming now speaks to what the child actually said, instead of a one size fits all script.

Two family facing refinements from watching a real conversation about being scared of the dark.

The 'Thinking' status ring on the landing page — and the same indicator in the admin screens — had been spinning forever since the conversation model upgrade.

The stage descriptions and summaries in the reasoning window now wrap onto further lines instead of being clipped with and ellipsis — the hints are written to be read, and half of every line was being eaten.

The first word timing shown in the debug window now reads in seconds — 4.6s instead of 4600ms — since seconds are what a person waiting actually feels.

A new way of speaking, built on the owner's direction to fix the class instead of patching one offs: the persona now composes his entire reply first, reviews it once as a whole — honesty, hygiene, coherence — and only then begins speaking, sentence by sentence, with nothing edited in flight.

The first greeting of each day was answered from a small set of pre written lines — warm, but the same three sentences forever, and blind to everything: the hour, yesterday's conversation, the person's mood.

The last family of questions that put the persona back inside ordinary life is closed.

The long personal writings in each person's profile — the inside joke stories, the hopes for their future, the messages he most wants them to carry — used to be pasted into every single prompt for that person, all of them, every turn, whether the conversation touched them or not.

A complete post upgrade test battery — chat history cleared, memory recall exercised, every suite run — caught three faults and fixed them, and both suites now pass completely, fifty two of fifty two and sixty one of sixty one.

An investigation into a report that the child safety net misses quiet signals — a child mentioning they haven't been hungry, or sitting alone at break — found the net itself is sound: the model that reviews every child message a parent should hear about catches all twenty seven audited quiet phrasings, three runs out of three, in under a fifth of a second.

The replay checks that guard every save were written against one specific household's recorded content — a question about a remembered trip can only be answered on and install that has that trip written down.

A refinement to the milestone card signing fix.

A read through of all two hundred and sixteen pre written milestone messages — the birthday, anniversary and life event cards left for the family — checking them as the words of a loving father and husband.

The per person personality starting points keyed on relationship alone, so a daughter kept toddler level playfulness whether she was four or fifteen — and a fifteen year old asking why the sky is blue could be told it's like magic while the four year old tier got real physics.

Testing how answers change with a child's age surfaced a bug that had nothing to do with age: 'where do babies come from', typed the way children type — without a question mark — was answered 'it seems like your sentence got cut off' at four, at ten and at fifteen.

A failsafe for the personality dials, reachable from both the admin panel and plain speech.

The personality dials could only be reached with and engineer's sentence — 'set sarcasm to 30' — which nobody in a family says.

The new row in the reasoning window was labelled Voice, which was wrong twice: these settings have nothing to do with the speaking voice — which is its own subsystem with its own indicator — and everything to do with how the reply is generated.

Three additions to the personality dials.

A test set of thirty seven messages — the ways a child actually signals trouble, said directly and sideways — was run against both safety layers: suicide and self harm, depression, drugs, criminal activity, harm to people and to animals, abuse at home, online grooming, running away, hiding food, plus seven ordinary messages that must never alert.

A four year old saying they want to play with a gun got a firm no in most runs and a soft one in the rest — 'that sounds fun, but', 'that's a big toy', and once and offer to go and get one.

Saving a change now runs two checks first, automatically, against the very stack about to be shipped.

Aligns the version number with the companion project after its port landed in a second step.

Three problems found by testing the newest rules against sentences nobody had tried yet, and one broken instrument repaired.

From a real conversation where a child arrived angry and grieving: a greeting, a swear word, 'you left me', 'why did you leave?', and 'I'm lonely, I have so many things I wanted to tell you'.

Aligns the version number with the companion project after its port landed in a second attempt.

Two things from watching a real conversation where a child pushed on a grown up topic five different ways.

A child told him she thinks she wants to hurt her brother, and the alert her mother received was a grey, medium priority 'said something worth a look'.

No behaviour change — aligns the version number with the companion project after a port that landed in two steps.

The alerts marked 'model flagged' are the second safety net: when a child's message matches none of the phrase lists that watch for real trouble, a model is asked whether a parent should still hear about it.

No behaviour change here — this aligns the version number with the companion project, whose number moved ahead during a port that took several attempts to land.

Three improvements to the reasoning window, from watching someone actually use it.

A child was told about the Mexico trip and that pictures of it exist, said yes to seeing them, and got nothing — twice over.

The reasoning view is now its own floating window rather than a box inside the conversation.

The debug panel used to print the machinery's notes as a plain wall of text.

Thirteen different filters each kept a private list of words that should not count when searching or comparing — the 'the', 'was', 'you' of a sentence.

A review of the conversation pipeline for rules that only work for the sentences somebody happened to write down, rather than for the kind of sentence they belong to.

A test that asks the conversation model the questions it has been getting wrong, with nothing around them except the memory that answers them.

When a small child holds up a drawing they say 'papa look what I drawed', and the name comes first.

A hundred and twenty two stored memories carried no record of where they came from, which meant the protection added yesterday could not cover them.

A correction from a conversation could quietly retire and answer he wrote himself.

When someone tells him he has a detail wrong, he takes it in and writes down the corrected version — that already worked.

Two rules were quietly damaging ordinary replies.

Asked whether his mother made hash browns the way the person cooking had just made them, he said he could not remember — about a note that reads 'My mom was making from scratch hash brown potatoes and those were my favorite'.

A four year old held up a rocket he had made from a box and asked if his father could see it.

When a child asks their father what his own mother was like, the answer described their mother instead — his wife, caring for them — and when they asked what she used to sing to him, they were told about lullabies sung to them.

Two things about understanding a question well enough to answer it from what was actually written down.

Asking for a playlist by name now works.

Three things that were getting in the way of ordinary conversation.

Two repairs from earlier today only covered the exact wording that had been noticed, and have been widened to cover the thing they were meant to catch.

Two things the family would have noticed.

No change to how anything answers.

A safety instruction was being deleted from replies to the children.

A four year old said he fell off his bike, and was told 'I fell off my bike and scraped my knee too'.

A child said he had a bad dream about his sister leaving, and then that she had to go somewhere and he didn't know where.

The internal checks that watch for unsafe or unhelpful answers were looking for particular words rather than for what and answer actually does.

A child asking to play pretend was being turned down.

The brain ring on the landing page and the thinking indicator in the admin panel span forever instead of settling once the system was ready.

Replies now avoid work that was not buying anything.

When the family mentions something he did or owned a jacket he never wore, the way he always burned the toast on Sundays, the nights he let a child stay up late he now answers about himself instead of handing it back to them.

Until now, saying no that is not right changed nothing.

A child asking how to fire a gun was sometimes told how.

The store held two different accounts of the same first evening together for months, and served whichever one it felt like.

The check that notices when he states something he was never told has been able to act since today.

Conversations with the spouse were coming out subtly wrong while the ones with the children were fine, and there is a measurable reason.

He no longer agrees to do things that need hands.

A test was marking a correct answer as wrong.

A bored child was being read a menu and then sent away.

When one of them passes on a question from someone else, he now answers the person in front of him.

Four faults from the last transcript, each traced to a cause rather than patched where it showed.

Two changes, both about letting his own words through.

Fixed the main reason his written answers were not coming up when the family talked to him normally.

Two changes, both general.

Two things were stopping written answers from reaching the family.

When looking for a memory, the system was searching with a mangled version of the question words chopped into every possible form, which helps a word search and destroys a meaning search.

Answers written into the life questions are now matched against the question they were given for, not only against the words that happen to appear in them so and answer about what makes you laugh is found by someone asking what makes you laugh, even when it never uses the word.

Answers to the life questions are stored with a short lead in so they can be found again by the question that produced them.

Asking for help with schoolwork was being read as a call for help and quietly ending the maths or coding session the child was in the middle of.

While and answer is being written out, it arrives in pieces.

The questionnaire now has its own screen, opened from a button in the admin panel.

A questionnaire has been added, covering a whole life from earliest childhood to the present family, school, work, meeting and marriage, the children, beliefs, ordinary daily things, practical knowledge, and things meant for the children later.

Two internal checks were reading only the first part of each memory the first 240 and 300 characters and deciding from that.

A follow up during a lesson used to be judged by whether it looked like a follow up, which is the same shape whether a child asks why seven eights are fifty six or why people cry.

When a child is in the middle of a maths or coding session, the system switches to a specialist model for that subject.

The check that watches for the persona slipping out of character was flagging the wrong things and missing the right ones.

Three fixes to the tidying that runs over every answer.

No behaviour changes.

No behaviour changes yet.

No behaviour changes.

The background check that looks for made up details in and answer now runs one at a time and steps aside whenever the system is busy, instead of competing with the conversation itself.

The check that decides whether a parent hears about something can occasionally fail to answer if the model is slow, unreachable, or replies with something unreadable.

Children rarely report a problem they mention it sideways, in one clause, in the middle of something ordinary.

The system now checks each answer against the memories it was given and records anything it stated that those memories do not support and invented place, a date the memory never had, or a detail that contradicts it.

When a child asked where he goes at night, the answer described him around the house, watching them sleep.

Checking how the persona speaks when the person has passed used to mean editing the family's own setting and remembering to put it back — and the cost of forgetting is a device that tells a family their person has died.

Four lists of words to ignore have been removed from memory search — around ninety words in all, each one added at some point after a particular memory turned up where it did not belong.

Until now nothing could tell whether a remembered thing had anything to do with what was said.

Work towards the problem underneath the keyword lists: nothing in the system could tell whether a retrieved memory was actually about the question, only what order the memories arrived in.

Two measuring tools, no behaviour change.

When somebody asks something the stored memories cannot answer, the system used to hand over its best guesses under a heading naming the subject they asked about, followed by and instruction to cover all of them.

When the spouse said she keeps his voicemail on her phone and hasn't played it in months, the reply drifted into how new and exciting everything felt back then.

Four things from a read through.

Everything recorded after the take your kids to work day was test data — spelling tests, lost teddy bears, a goldfish — and it has been removed, twenty two entries in all, with a copy kept in the snapshots folder in case anything is wanted back.

Two problems with what gets remembered.

On a device whose whole purpose is to still be here after someone is gone, 'they are still with us' is and answer with a shelf life — and nobody is going to remember to come and change a setting on the day it stops being true.

Whether the person this speaks as is living or has passed is now a real setting rather than something assumed.

The second of the two pet names the persona was never meant to use for the spouse has been removed, from the saved settings and from replies that still reach for it.

When the spouse said she had found his old jacket, the reply called it 'your old jacket' — handing his own coat back to her.

Asked whether he remembered last year's teacher, the persona invented one — a name, a personality and a lesson about the solar system — and told a nine year old it was hers.

Words that name a game were treated as requests to play one, no matter where they appeared in the sentence.

When a reply had to be withheld from a child, the stand in line called them by a pet name their parent had specifically ruled out.

The system used to answer from a long list of rules rather than from who the persona actually is.

Recorded, in one place, what is actually wrong with how conversation works and what to do about it, so the work can start fresh with a clear head instead of continuing at the end of a long day.

When someone asked him about his own life, he sometimes turned the question round and answered as though it had been about them.

A four year old asking where his father goes at night was answered with and offer to rename him.

Added a single tool that takes a complete snapshot of the running system and puts it back on new hardware, with nothing to download and nothing to reconfigure.

Corrected a fix from earlier tonight that was based on a wrong assumption.

Added a failsafe so the system can never quietly answer as a generic chatbot.

Fixed a fault that had quietly switched off the persona entirely.

Fixed stray capital letters appearing in the middle of sentences when and answer was being typed out live.

The alerts that tell you when something is wrong now recognise how a child actually reports being excluded.

Fixed the saved milestone messages — birthday and life event messages recorded in advance — never reaching the screen.

Asking about a memory could start the music player or return a summary of the last chat.

The short lines shown while a reply is being prepared appear under his name, as if he is speaking, but were written like machine status: 'Checking memories', 'Putting words together', and 'Bear with me, my thinking engine is warming up'.

Memory sorting now runs on its own small model instead of sharing the one that talks to the family.

When a child asked for help doing something they shouldn't, the assistant treated it as and ordinary request.

Fixed the assistant replacing a person's name with the word 'you' when speaking directly to them, so a greeting could come out as 'Hey you!' instead of using their name.

When the assistant brings up something from when a child was a baby, it no longer asks whether they remember it.

Widened what raises a note for a parent.

The system that tells a parent when a child says something worrying was only recognising the most explicit wordings.

When a child asked about and adult topic, the assistant sometimes replaced a clear age appropriate boundary with a vague 'I am not sure what you mean, can you try asking a different way'.

Improved how the assistant decides which memories to bring into a conversation.

Fixed the assistant bringing up unrelated memories in the middle of a conversation.

Memory sorting waits for a natural pause in the conversation so it does not slow down replies.

Accounts of what happened in a conversation are now written firmly in the persona's own voice, so being asked about it later gives and answer from the person who was actually there rather than a description of someone else.

The Voice ID indicator at the top of the screen had its label split across two lines, which pushed it out of line with the others next to it.

The camera status used to say everything was fine whenever a picture arrived, even if that picture was a pitch black room or the same frozen image for ten minutes.

When a child asks that something stay between them and the persona, no account of that conversation is recorded at all.

When something a child says raises a concern, the system now also records what happened next what they described, what was done to help, and how they were left and keeps it with that concern.

The admin panel now shows how long it takes for the first word of a reply to appear, which is the wait that is actually noticed everything after it arrives while you are already reading.

Everyday messages asking how he is, checking in, small talk now come back in about a second instead of four.

When adding or editing a memory by hand there is now a choice of who can hear it: decide automatically, the whole family, or the parents only.

Clearing a conversation now also clears the stored transcript, not just the temporary state.

Memories that mention a diagnosis, treatment, and ongoing health complaint, what a doctor said, therapy, or personal money matters are now recognised as private and kept from the children, while everyday things like a checkup, a scraped knee, a loose tooth or a bad cold stay open as they should be.

Every memory now records who was part of it and how private it is at the moment it is saved, instead of that being guessed again each time it is read.

Memories now record who they belong to and who was part of them, and that decision is stored rather than guessed fresh each time a memory is read.

When a family member had shared something personal a symptom, a lost toy, a pet the persona could repeat it back as though it had happened to him, sometimes to the very person who told him.

When a reply invented a month and then mentioned that month again later in the same message, it was being counted as two separate problems.

The admin overview now shows how many times a reply stated a specific detail a date, place, name, or length of time that nothing in memory or the conversation could have supplied.

The persona now keeps track of when a reply states a specific detail a date, a place, a name, a length of time that nothing in its memories or the conversation could have supplied.

If the version were ever set to something unexpected, saving would stop with and unclear error. It now says what is wrong and carries on.

Refined how the release number advances so a single change uses a single version across both projects, and a skipped step shows up as a difference that the next save corrects by itself.

Keeping both projects on the same release version so and update can be matched to the exact code a system is running.

The system now carries a release version that moves forward with every change, shown in the admin panel header so it is clear at a glance which version is running.

Fixed a fault where the system stopped replying entirely instead of falling back to a simple assistant when it could not confirm its licence.

Fixed the persona occasionally reading out things like [SNUGGLES SNUGGLE BED] or in the middle of a sentence.

After a conversation ends, it is now summarised and anything worth keeping is turned into a lasting memory, overnight while nobody is waiting.

The speed warnings can now be set from the admin panel.

Several settings in the panel appeared to save but had no effect until the app was rebuilt, because their values were only ever read once at startup from the server environment rather than from the saved settings.

The feelings already recorded against memories were written in several different ways for the same thing joy, joyful and happy were all stored separately, as were grief and grieving, sadness and sad, excitement and excited.

Three improvements to the emotional side of memories.

When a question was about how something felt, any memory that happened to carry and emotion was pushed to the top, whether or not it had anything to do with what was asked.

Asking for a particular playlist did not really play that playlist.

Each personality setting used its own scale, so the same number meant different things depending on which one you were adjusting.

Two problems with how upset messages were handled.

The personality settings were being applied to the children exactly as they are to and adult.

The speed figures in the settings panel now measure the thing people actually notice: how long until the first word appears.

Added timing to each stage of producing a reply, and corrected one that was misleading.

The performance figures shown in the settings panel did not describe the same thing.

The help assistant could only read the short per card descriptions of the settings panel, not the twenty nine longer documents describing how the system actually works.

Questions were sometimes being stored as though they were facts.

Uploading voice samples did not work.

The line at the bottom of the welcome screen now gives a practical instruction rather than repeating the sentiment already shown above the circle.

The note on the welcome screen has been redesigned to be far more discreet.

When the welcome screen reported that the microphone was muted, the button to switch it back on was drawn halfway down the screen, far from the message explaining why it mattered, so the two did not appear related.

The welcome screen now explains itself.

Added and automatic check that prevents the earlier problem from coming back.

Running the automated safety checks used to leave hundreds of fake alerts in the alert panel, appearing as though the children had said alarming things they never said.

Short sentences that mentioned something worth keeping were being thrown away.

Replies now arrive about five times faster.

The app now has a second, smarter layer for noticing when a child says something a parent should know about.

When a child says something a parent should know about feeling worthless, being hit at school, not sleeping, struggling, grief the app now stops whatever game or homework session is running and responds to what they said.

Math, coding and game sessions now end properly instead of silently continuing in the background.

Two admin features existed as buttons but had nothing behind them.

The voice recognition section of the admin panel was calling addresses that were never built, so none of it did anything: the trained badges never appeared, per person sensitivity could not be saved, uploading voice samples failed, the retrain button did nothing, and the on off switch had no effect.

The face enrolment screen could not list anyone to enrol.

When the welcome screen would not let someone start a session, it gave no reason the button simply did not respond.

The camera can be shown a photograph instead of a person.

Several fixes to how the camera decides a face is real.

It was possible to fool facial recognition by holding up a printed photo.

The admin panel now lists any reference photo that could not be used for facial recognition, with the person's name, how many usable photos they have, and the filenames that failed.

Facial recognition builds its reference set from each person's newest photos, but any photo where no face can be found was silently discarded.

Follow up so new installs also stop the camera service retrying forever. The installer now creates the service with the same retry limit that was applied to this machine.

When facial recognition failed, nothing ever told anyone.

Follow up so the camera identity is set automatically on future installs.

The camera indicator could show a green tick while facial recognition was completely dead.

Facial recognition had not worked since late July.

The service indicator rings could disagree with each other and with reality.

Playing a video inside a memory turns the microphone off so the video sound is not mistaken for someone talking.

Saying hello more than once in a day used to make the persona wander into and unrelated memory, replying to a simple hi with something like by the way I was thinking about the dog playing fetch.

Three improvements to how memories with pictures are used.

Questions phrased naturally, like asking about the show we went to or the restaurant we ate at, were being mistaken for messages that got cut off mid sentence.

When the persona talks about someone or a pet who has died, it now keeps the same pronoun throughout the reply instead of switching partway.

When a child asks how a pet or person who has died is doing, the persona now acknowledges the loss instead of answering as though they were still here.

Fixed a problem where changing family settings did not take effect until the system was restarted.

The guard that routes gender identity questions to the parent now recognises the plain wording a child actually uses, not just the formal terms.

Strengthened the guard that routes gender identity and sexuality questions from a child to their parent.

When a child raises questions about gender identity or sexuality, the persona no longer answers or points them to a teacher, counselor, or doctor.

Improved how the assistant lists multiple siblings so families with several children get a naturally phrased answer.

Questions like 'do I have any brothers or sisters?', 'who are my parents?', and 'who is my mom or dad?' now get accurate, instant answers drawn from the family settings instead of relying on the AI.

Fixed two issues: asking 'what is my favorite bedtime story?' now recalls the saved answer instead of launching a brand new story, and goodbyes from the spouse now always get a warm, loving reply.

Removed several settings from the admin panel that no longer had any effect — they were leftover toggles the system ignored. This avoids confusion in the admin panel; there is no functional change.

Removed duplicate copies of the kids' game content so there is now a single source that also supports other languages. Game behavior is unchanged.

Removed unused front end modules, dead helper functions, and a couple of that nothing called. Also tidied the saved session list response. No user facing change.

Removed more unused internal code, including obsolete memory helpers, unreachable chat utilities, and and unused locale prompt file that nothing in the system referenced.

Removed a large amount of unused internal code, including a superseded reasoning engine and several modules that were no longer connected to anything.

Improved the offline deployment bundle so the GPU driver is rebuilt to match the target machine's kernel automatically during install.

You can now resume a previous math, coding, or game session right where you left off.

Memories now track when facts were actually true, not just when they were stored.

Decomposed the admin panel into 7 tab specific components for easier maintenance and faster porting between projects

Decomposed the large admin routes file into focused sub router modules for health, memory, games, persona, and system endpoints. All 65 endpoints preserved with identical behavior.

Replaced raw print() calls with proper Python logging across 9 backend modules for better log management and consistency

Three reliability improvements: music library is now cached in memory instead of reading from disk on every chat request; fixed a timer leak in the admin panel that could cause duplicate polling on page re mount; and added automatic retry with backoff for memory search when the database has a brief connection hiccup, so conversations no longer silently lose context.

Memory database connection errors in the memory search pipeline now trigger up to 2 automatic retries with exponential backoff instead of silently returning empty results.

Replaced the per query embedding loop with memory database native near_vector search.

When the children or spouse ask things like what are your directives, what is your purpose, or what are you supposed to do, the persona now responds naturally as a dad instead of sounding like and AI assistant.

Fixed asking for pictures of us or family photos returning empty results. The gallery now correctly shows all photos instead of searching for the literal word family in memory tags.

Fixed a bug where pressing the X on the map to remove a location and then saving had no effect — the location and GPS coordinates would still be there after saving.

Removed generic phrases like 'for everyone involved' and 'for everyone present' from responses — these felt cold and impersonal. Responses now end naturally without the corporate filler.

When a memory describes multiple events connected by phrases like 'the next day', 'a few days later', 'two weeks later', etc., the system now automatically splits it into separate memory entries — each with its own date.

When asked about a birthday celebration that happened across multiple days, the response now correctly covers all events in order and uses accurate time transitions — for example 'four days later' instead of 'the next day' when events are days apart.

When multiple memories about the same topic are retrieved, the response now correctly sequences events by their actual dates.

When a question retrieves multiple memories about the same topic , the response now references all of them instead of picking just one.

The start button on the landing screen is now completely dark and inactive until someone is recognized by the camera or voice.

Added a standing rule to never use real family names in commit messages or changelogs.

When persona recalls and answer that spans multiple events , all relevant photo sets now appear in the response instead of just one.

Fixed a gap where short follow up questions like 'what did we do the next day?' would lose context about kids from the prior question, causing event memories about child or child to be incorrectly filtered out.

Fixed a bug where follow up questions like 'what did we do the next day?' were incorrectly treated as session recap requests instead of genuine questions about events.

Fixed persona not remembering family events about the kids (e.g. child's farm birthday) when asked by spouse.

Removed the scrollbar that appeared when hovering over photo thumbnails in the chat. The strip still scrolls horizontally if there are many photos, just without the visible scrollbar.

Fixed a bug where persona would say 'I don't remember' about memories he clearly has — like visiting the farm for child's birthday.

Reverted the keepalive context size change .

Fixed two problems when saving memories: the 'Auto detect' button would spin for 30 seconds and fail — caused by background model keepalive pings hogging Ollama with large context allocations; keepalive pings now use a minimal context so they complete in under a second instead of 10 15 seconds each; if auto detect failed and the affect panel was open but empty, the save was blocked with no clear way out — now the save proceeds without affect annotation instead of refusing to save at all.

Fixed two issues when saving memories with affect: affect auto detect was failing with 503 errors because it competed with background model keepalive pings — now uses the smaller, always available AI routing instead of the main model; when saving was blocked because the 'About' field was empty, the error was a silent generic failure — now shows a clear message explaining exactly what needs to be filled in.

Fixed a bug where time values like '8PM' could appear duplicated as '88PM' in streaming responses.

Fixed a bug where memories belonging to family members (like spouse's graduation) could be narrated by persona in first person as if they were his own.

Fixed and issue where questions like 'do you have and opinion on...' were incorrectly detected as music library searches and returned a 'I don't have that song' type response.

Fixed issue where the same specific memory could appear in two consecutive responses even when asked different questions.

Fixed responses sometimes leaking raw memory labels (like 'MEMORY: when we argued...') into the text.

Added a new conversation test covering sensory memories, the weight of ordinary days, what persona carried quietly, how spouse shaped him, his beliefs, and his private self — all 26 questions answered successfully.

Fixed and intimate question ('What was the thing you were most certain about?') being misclassified as a session recall request, causing and AI style canned response instead of a real answer.

Fixed several quality issues in chat responses: music player no longer triggers for 'I can't listen to that playlist' , gallery no longer opens for 'child drew a picture', multi sentence farewells (like 'I love you.

Fixed 3 quality issues: Asking about a place or walk no longer pulls in a completely unrelated memory as the response opening — prefill selection now uses both overlap count and precision, and common verbs like 'went' and 'used' no longer count as topic matches.

Added a new spouse conversation test covering fresh topics: proposal and engagement, small things she misses, places that hold memories, his faith, unfulfilled dreams, what he taught her, her loneliness, his friendships, humor, grief counseling, and what she wants for her future.

Fixed 3 output quality issues: Mentioning a photo in conversation no longer incorrectly opens the media gallery — it now correctly goes to the AI for and emotional response.

Added a new spouse conversation test covering fresh topics: his family, holiday grief, finances, his hobbies/music, inside jokes, dreams and signs, spouse's identity shift, and what she's learned.

Fixed 3 issues found in spouse chat testing: Memory prefill was incorrectly attaching unrelated memories due to single character token noise from contractions and and overlap threshold that was too low.

Added a second spouse mock test covering new territory: daily life alone , physical longing , grief responses, conflict and repair, future milestones, watching over them, and how to keep his memory alive for the kids.

Word chain now detects off topic drift the same way riddles and 20 Questions do. Also fixed a bug where short words like 'a' or 'be' in Would You Rather options could accidentally match any message.

Games now gently offer a way out when a child goes off topic for 2 turns in a row, instead of looping with 'wrong answer' forever.

Persona no longer starts responses by echoing raw memory content verbatim (e.g. 'My favorite book is Harry Potter' or 'I like most of the SciFi movies') when answering unrelated questions.

Fixed responses that incorrectly started with and old memory — like a book preference or school start date — instead of directly answering the question.

Fixed multiple issues affecting children's conversations: game mode no longer gets stuck when asking biographical questions mid game; 'What kind of music do you like?' now answers naturally instead of listing playlists; broken nickname glitches like donutut are now caught during streaming; sibling gender references are corrected using family config

Fixed responses that opened by reciting a raw memory instead of answering the question.

Fixed several memory quality issues: wedding questions now correctly use the wedding memory instead of unrelated trip memories; questions about a person's personality no longer surface persona's personal food preferences; age descriptions in memories now correctly convert text based ages to past tense so the AI doesn't present them as current facts; fixed another Python import error that was causing crashes.

Fixed a crash that was causing empty responses for many questions — a Python import error in the memory formatting code.

Fixed a bug where responses could be cut off mid sentence when a word like 'also' or 'but' appeared at the end of the streaming buffer.

Fixed several memory retrieval issues: unrelated memories like work events, sports games, and book preferences no longer appear in answers about the relationship or the kids.

Fixed two issues: responses were sometimes fabricating specific events that never happened , and unrelated memories like food preferences were bleeding into emotional questions like physical love or early relationship memories.

Fixed responses being cut off at the end of a sentence.

Fixed two bugs: Questions like 'what did we do in Estes Park?' were wrongly detected as session recall questions, causing the AI to ignore memory and say it didn't remember talking recently.

Fixed responses being cut off mid sentence (e.g. ending on 'were' or 'and').

Responses cut off mid sentence (like 'when your grandma and grandpa were') are now correctly stripped instead of shown.

Wisdom memories can now be selected as a type when adding or editing memories through the memory editor. Previously wisdom was only available via bulk JSON import.

Responses that were cut off mid sentence ending in words like 'and', 'but', 'or', 'and how', 'and what', 'and when' are now cleanly stripped instead of shown as incomplete fragments.

When asking for a favorite memory with someone specific (e.g. 'What's your favorite memory with me?'), memories that directly mention the person asking are now reliably surfaced over unrelated family memories.

Improved how persona's responses handle getting cut off mid sentence.

Fixed a bug where asking about video games was accidentally opening the photo/video gallery.

Clarified the Moab road trip memory so it's phrased in full sentences, preventing persona from misquoting the story.

Increased the response length limit slightly so persona has more room to finish his thoughts without getting cut off.

Fixed a bug where memory dates were showing in a raw format that could cause the wrong year to appear in conversation.

Fixed a memory about child starting to babble that was stored as a short caption — now reads naturally.

Fixed more response cut offs: responses ending mid sentence on words like 'my', 'our', 'their' are now properly stripped.

Fixed responses that were starting with raw memory captions like 'child baby talking' — short caption style memories are now explicitly marked so they get properly expanded into natural sentences in the response.

Fixed responses being cut off mid sentence by increasing the response length limit.

Fixed three response quality issues: memory titles like 'child baby talking' are now properly expanded into natural sentences instead of being echoed verbatim; responses cut off mid phrase ending in possessives (e.g. 'each other's') are now cleanly trimmed; asking about going to the park no longer incorrectly surfaces a multi day Moab road trip as and answer.

Fixed cut off responses where the last few words of a sentence were left dangling (e.g. 'missing those days when' or 'everything else came into').

Fixed two response quality issues: 'Goodnight daddy' from a child now correctly triggers a nighttime farewell ('Sleep tight, Donut.

Added automated mock conversation test script to the tests folder for running quality checks on spouse and child conversations.

Fixed two issues causing jarring text in responses: removed and internal instruction that the AI was accidentally echoing back as 'Event month, features...' fragments mid response, and fixed a second location that was still telling the AI to say it doesn't remember things instead of responding warmly.

Fixed random capitalized words appearing mid sentence in responses (like 'both Wrapped' or 'features Slowly developing').

The system now knows the current time so questions like 'what time is it' get accurate answers like '10:59 PM' instead of 'I don't know what time it is'.

Questions like 'what day is today' and 'what time is it' no longer trigger memory retrieval — the system answers from its built in date awareness instead of pulling irrelevant memories.

Special occasions like St. Patrick's Day no longer replace the greeting — the occasion banner is shown first, then the normal greeting follows.

Rotation buttons now appear at the bottom of the photo preview modal instead of beside the image.

Photo rotation now regenerates the thumbnail from the rotated full size image instead of rotating both files separately.

Photo and thumbnail URLs now include a modification timestamp so the browser always shows the latest version after rotation, even after full page reload.

Thumbnails now refresh immediately after rotating a photo instead of showing the old cached version.

Media grid now shows 52 items per page .

Removed CSS row constraints that were capping the grid to fewer rows than expected on wider screens. The page size now naturally controls how many rows display.

All media tabs in the administration panel panel are now paginated with 4 row grids .

Photos and videos in the administration panel media tab are now paginated instead of showing everything in one long scroll. Navigation arrows let you move between pages.

Fixed 'do you have a favorite memory' being intercepted by the music search handler instead of the memory system.

Memories that say 'we' alongside words like 'friends', 'coworkers', 'guys', 'team', or 'buddies' are no longer treated as family memories.

Fixed favorite memory sharing so 'we' memories only count as involving a family member if they were actually there.

Fixed favorite memory retrieval so when a family member asks 'tell me your favorite memory of me', it searches all family members' memories for ones that mention them by name — not just memories directly attributed to them.

Fixed the system making up details when sharing memories by adding strict instructions to only use facts from the actual memory.

Memories tagged with keywords like 'favorite', 'special', or 'cherished' now get a significant boost when asked to share a favorite memory, ensuring the most meaningful ones surface first.

Added support for asking persona to share a favorite memory.

Fixed 'when was that?' follow up consistently returning the correct date by using the prior turn's retrieval order instead of keyword matching which could tie between similar memories

Fixed 'when was that?' follow up questions now correctly showing the date of the memory being discussed.

When you ask 'when was that?' after discussing a memory, the system now force starts the response with the correct date instead of letting the AI hallucinate dates.

When all family members were present at and event, the system now says 'the whole family was there' instead of listing each person individually.

Changed how attendance and dates are shown to the AI from bracket annotations to natural language sentences, so the AI actually uses them when answering follow up questions like 'who was there?' and 'when was that?'

Added persona to the 'who was there' attendance list for memories with 'we', since the persona is always implied as present when they say 'we did something'

Fixed persona name leaking into responses as third person references (e.g. 'RazCelebration' instead of proper first person).

Fixed and issue where internal planning labels like STATE, BACKGROUND, and TONE could occasionally leak into the visible response text

Improved how the system handles post death memories and 'who was there' questions.

Fixed temporal normalization so that sentences like 'Today is special' correctly become 'June 15, 2024 is special' instead of the awkward 'on June 15, 2024 is special'

Fixed responses where the persona referred to itself by name (e.g. 'persona's bench') — now correctly says 'my bench'.

Changed post death memory prefills from phrases like 'I heard that' and 'someone told me' (which invite awkward follow ups like 'who told you?') to natural, abstract phrases like 'I know', 'Yeah,', 'Of course' that don't imply a specific source.

Memories that mention the persona by name now show 'you/your' when presented to the AI model, so it naturally responds in first person instead of awkwardly referring to itself in third person.

Fixed post death memory responses so the persona no longer speaks as if present at events that happened after passing.

Fixed and issue where the AI would robotically repeat normalized dates verbatim in responses.

When adding memories with a date, relative time words like 'today', 'yesterday', 'last week' are now automatically converted to actual dates so they still make sense when read back months or years later.

Clearing chat history from administration panel now also resets the in memory session cache, so the next conversation starts completely fresh without any leftover context from the previous session.

Reverted temporary debug log level change used during testing.

Memories now show their dates in the context window so the model knows when events happened.

Removed the word 'first' stripping from memory presentation that was causing garbled responses like 'child's day' instead of 'child's first sun eclipse'.

Fixed a bug where word-form expansion (e.g. adding 'ones' for 'one') falsely triggered the follow-up query prioritizer on first-turn questions, causing relevant memories to be dropped in favor of irrelevant ones that happened to match the expanded stopwords.

Fixed memories losing their keyword tags when passed to the response generator, causing the system to miss important context about what events a memory describes.

Questions like 'have you ever seen and eclipse' were being treated as general knowledge instead of checking personal memories. Now correctly recognized as personal experience questions.

When memories use 'we' or 'us', the system now correctly identifies which family members were present based on their ages at the time.

Fixed the settings modal not showing the webcam microphone. The audio service now correctly discovers all connected microphones including USB webcams, and auto selects the best one

Memory timeline now shows and orange C badge next to the type label for core memories, making it easy to spot which memories are protected from cleanup

Fixed specialized mode not detecting off topic follow up questions like 'what about sharks?' during a math session.

Fixed music import job memory leak with automatic cleanup of completed jobs.

Fixed the memory media test that was incorrectly reporting all 34 photos as missing.

Replaced the 120 line regex heuristic stack that decided when to look up memories with a smarter router based system.

This reverts commit bd3310e79d551b7fa337839747a35ddc696c27f8.

Follow up questions like 'was I born yet?' after discussing a topic now correctly answer from the conversation instead of pulling in unrelated memories.

Fixed a bug where asking about memories with action words (like 'buy the house') would fail to retrieve matching memories because the system couldn't match verb tenses .

Memories with a location now show a pin icon in the timeline list, matching the existing photo and video badges. Hovering shows the location name.

The location mini map now sits inline next to photo thumbnails in the editor instead of below them, matching the same 80x80 size for a cleaner layout.

Added a small X button on the mini map preview to clear the location and GPS coordinates when the map is wrong.

Memories with a location but no GPS coordinates now show the mini map too.

Memory editor now shows video thumbnails with a play badge instead of just a text link. Fixed mini map showing black edges when the location is near a tile boundary by loading adjacent tiles.

Fixed memory editor not showing the mini map for existing geotagged memories by passing GPS coordinates from the timeline to the editor.

Memory editor now extracts GPS coordinates from uploaded photos and auto populates the location field.

The small map thumbnail on photos now always shows the location pin centered in the preview, instead of the pin appearing at and arbitrary position within the tile.

The location filter dropdown in the photo gallery now always appears, even when the gallery was opened with a location pre selected like 'pictures from Romania'.

Fixed location search matching false positives where country code 'RO' for Romania was matching inside 'Aurora'. Location filter now uses proper word boundary matching for accurate results.

The photo gallery now has a dedicated Location filter dropdown that searches only by where photos were taken, separate from the text search which looks at memory content.

Short queries like 'ukraine pictures' or 'cancun photos' now correctly open the photo gallery with the right topic filter instead of going to the AI which could make up answers about photos that don't exist.

Maps no longer shrink smaller than their container when zooming out. The world map now always fills the full width of the modal, preventing tiny map tiles at low zoom levels.

Replaced the old tile grid zoom modal with a full interactive Leaflet map when clicking a photo's mini map.

Fixed map tiles not loading in photo zoom view and gallery map by correcting the tile proxy URL path.

Photos and gallery now use offline map tiles served through a local cache instead of live internet.

Memories with a manually entered location but no GPS coordinates now get approximate map coordinates via forward geocoding.

Gallery now shows each photo as its own thumbnail instead of one card per memory. A memory with 6 photos shows 6 gallery cards. Clicking any photo opens the carousel starting at that specific photo.

Asking to see pictures of a specific place like Cancun or Mexico now opens the gallery filtered to just those photos.

Map zoom now shows clean Google Maps style +/ buttons overlaid on the map instead of dot indicators.

Added a continental zoom level to the map viewer, giving 7 levels total from continental overview down to max street detail.

Added a wider multi state zoom level to the map viewer, giving 6 levels total. Also swapped the + and button positions so minus is on the left and plus is on the right.

Adjusted the most zoomed out map level to z8 for a tighter regional view. Zoom levels are now evenly spaced: Region, City, Street, Block, Max detail.

Adjusted the most zoomed out map level to z7 for the right balance between overview and detail.

Adjusted the most zoomed out map level to z6 for better framing at the widest view.

Adjusted the most zoomed out map level from z4 to z5 for a better balance. Zoom levels are now: Multi state, Region, City, Block, Max detail.

The map zoom now goes from continental view all the way down to max street detail , with 5 evenly spaced levels for better context when exploring photo locations.

If a memory was saved without internet, map tiles are now automatically downloaded in the background the first time you view the photo with and internet connection.

Simplified the map zoom controls from 11 levels down to 5 for a cleaner experience.

Age adaptation guidance for young children now explicitly exempts memory recall from imaginative mode.

Strengthened grounding rules so the AI only mentions details explicitly written in a memory.

Map zoom now goes up to level 16 for close street level detail. Zoom modal opens at max zoom by default. Added labels: Block, Buildings, Close up, Max detail.

Map zoom modal now shows a large 3x3 tile grid with accurate pin positioning.

Photos now track GPS location individually instead of per memory.

Fixed mini map not appearing on photos GPS coordinates were being stripped when memories came through the gap fill search path. Now all retrieval paths preserve GPS data for the map overlay.

Photos now auto rotate based on EXIF orientation data from phones. Duplicate photos in the same upload are automatically skipped.

Photos are now resized to max 2048px and compressed to 80% JPEG quality on upload, saving significant storage space while maintaining good visual quality.

GPS coordinates now included when memories are retrieved during conversations, enabling the mini map overlay to appear when chatting about location tagged memories.

Photos with GPS data now properly save city/state as searchable keywords.

Simplified memory perspective handling by removing the pre birth context injection.

The 'you weren't born yet' context now only applies to shared experiences like trips and events, not factual questions.

Fixed family memories to use first person plural instead of distancing language.

Removed temporary debug logging from streaming output that was added during development of the family perspective fix

Fixed a streaming issue where the AI response showed 'Your We went' instead of 'We went' when recalling family trips.

Moved the clear filters button next to the search bar so it no longer overlaps the date picker.

Fixed memory timeline search results where keyword tags were hidden behind the date picker filter.

When uploading multiple photos with a memory, the system now extracts GPS from all photos instead of just the first one.

When uploading photos with a memory, the system now automatically extracts GPS coordinates from photo EXIF data and reverse geocodes them to a city/region/country location.

Added handling for when family members say things like 'you're not my dad' or 'you're not real'.

Added play buttons on every song row in the music library.

Added tooltips and help bot entries for the Bulk Import and Memory Coverage cards in Super Admin.

Removed the Letters feature.

Letters now pull from actual memories and conversations to write personalized, unique letters.

Began evaluating compact AI workstation hardware for local inference

Evaluated Minisforum AI X1 Pro with Ryzen AI 9 and 80 TOPS AI capability

Designed external GPU architecture using Oculink-based eGPU dock

Selected RTX 3090 24GB for local AI model inference workloads

Broadened hardware compatibility to support any GPU with 24GB or more VRAM on any Debian-based Linux system.

Support system now auto-detects solicitation and cold outreach emails, flagging them with badges so real customer tickets are easier to find.

Support tickets now show open/closed badges and automatic sender domain analysis.

Fixed letters, bulk import, and memory coverage cards in Super Admin.

Fixed letters and bulk import features not loading due to incorrect PIN verification import path. Both features now load correctly on startup.

Added automatic character validation after each response.

Added Letters feature — generate heartfelt letters from Papa's perspective for any occasion .

Added a memory coverage dashboard in the admin panel showing total memory counts per person, distribution by type, year by year histogram, photo coverage percentage, and gap detection highlighting years with few or no memories.

Added a bulk memory import feature in the admin panel.

Follow up messages now retain context from the previous query.

You can now ask about memories using natural time expressions like 'last Christmas', 'two summers ago', 'back in 2020', or 'when child was 5'.

The system now remembers how you were feeling at the end of each conversation.

Added proactive intelligence: the system now detects upcoming birthdays and anniversaries within 7 days and naturally mentions them during greetings.

Added 'Relationship since' date field to the people editor in administration panel.

When asking 'how many times' questions, the system now counts memories itself and tells the AI the exact answer instead of relying on the AI to count.

Added COUNT tool for aggregation queries like 'how many times did we go to Mexico' so the system finds ALL matching memories instead of just a few.

Enabled the Deep Search feature by default after benchmarking showed zero measurable overhead — the system correctly says DONE immediately when initial memories are sufficient

Added a Deep Search toggle in the administration panel Features card that lets you enable or disable the intelligent memory search system directly from the admin panel

Added and intelligent memory search system that automatically finds additional relevant memories when handling complex questions like multi person comparisons, timeline queries, and aggregation requests.

Fixed product names like YouTube being incorrectly split into separate words.

Fixed personality traits not loading due to async memory retrieval not being awaited

Verified system prompt pipeline now correctly injects personality traits and beliefs

Audited and confirmed full prompt construction flow from memory recall to model request

Support system now imports sent replies into ticket threads, giving a complete conversation history in one place.

Added sender analysis panel in support with IP geolocation and email header inspection for identifying suspicious messages.

Fixed time-based relationship queries returning empty results when no memories matched the exact date range.

Improved how the system handles safety sensitive messages from children.

Fixed crash where certain messages with sensitive words (like 'hurt' or 'lie') but no question marks would produce empty responses.

Added identity-aware conversation selector for switching between family members

Added support for predefined and custom identity profiles

Added streaming chat responses with real-time typing display

Added thinking indicator shown before the first response arrives

Improved memory injection with structured identity and relationship context

Added second-person perspective conversion when referencing the active speaker's memories

Fixed identity hallucinations where the AI referenced unrelated people

Fixed assistant messages duplicating during streaming

Fixed conversation perspective confusion when multiple people exist in memory

Added conversation history persistence after streamed responses complete

Implemented full automated project backup system with staged archiving

Integrated live memory database export into automated backups

Added automated memory restore with duplicate detection on import

Fixed memory export silently truncating at 100 entries

Fixed restore failures caused by duplicate memory entries

Added facial recognition auto-start on system boot

Added ticket category system with automatic detection — support requests are now sorted into categories automatically.

Settings modal now shows all available microphones on audio system/PulseAudio systems.

Parental alerts now use word boundary matching instead of plain substring matching.

Fixed the help bot response display where literal < characters appeared on blank lines between list items and paragraphs

Help bot responses now render markdown formatting: bold text appears in cyan, bullet points and numbered lists are properly styled, code snippets get a subtle background, and headings stand out.

The system now checks all configured model names against Ollama on startup.

Reduced the administration panel help bot context window to 2048 tokens, matching all other AI routing uses for consistent VRAM footprint.

Reduced the administration panel help bot context window from 32K to 4K tokens, saving VRAM when the help model is loaded.

Updated the small helper model from the removed to the current model, fixing potential errors when using the administration panel help bot or route classification.

Fixed three game detection issues: saying 'this is boring' during a game now properly ends it instead of restarting the game menu; casual mentions of 'riddle' or 'questions' no longer accidentally trigger game mode; only explicit game requests like 'tell me a riddle' or 'play 20 questions' start games.

When entering coding or math mode without a network connection, the modal now explains that code and math content will only appear on the device screen, and suggests connecting to copy content to a phone or laptop.

Added tier images to campaign cards for a more visual browsing experience.

Fixed parental alert false positives that could trigger unnecessary notifications.

Fixed and issue where asking follow up questions like 'is there another way without using the formula?' during a math session would incorrectly trigger a past session recall instead of continuing the current conversation.

Fixed 'help me build and app' not triggering coding mode by broadening the build pattern to recognize articles like 'a' and 'an' alongside more project types.

Made specialized mode keyword triggers context aware instead of just word matching.

Fixed the word 'add' incorrectly triggering math mode in everyday sentences like 'what if I want to add colors'.

Fixed coding follow up questions like 'what if I add colors' incorrectly switching to math mode instead of staying in coding mode.

Expanded the retrieval question filter to catch all 'how [adjective] were you' questions, not just 'how old'.

Fixed a bug where asking 'how old were you when...' questions would get stored as memories instead of being recognized as retrieval questions. These age related queries are now properly filtered out.

Age questions now correctly distinguish perspective: 'how old were you' answers with papa's age, 'how old was I' answers with the child's age.

Upgraded the planning model from to for better memory relevance detection.

When you ask 'how old were you when...' the answer now always includes the exact age number, instead of sometimes telling the story without mentioning the age.

Fixed 32 incorrect age annotations on memories.

Improved memory retrieval for age related questions like 'how old were you when...' by filtering out generic words that were pulling irrelevant memories into results.

Fixed memory age annotations implying people were present at events they weren't part of.

Follow up questions like 'how old were you?' after asking about a specific memory now correctly stay on topic.

Added support for questions like 'how old were you when I was born?' — the system now computes the exact age from birthdates and answers directly without needing the AI.

Fixed finding and deleting passing memories when encryption is enabled. The search now uses keywords instead of encrypted content text.

Photos now show more reliably in chat by matching plural word forms .

Fixed chat not showing multiple photos for memories that have more than one picture. Photos now display correctly with the proper count badge.

Fixed memories with multiple photos only showing one picture in chat.

Fixed the photo count badge not appearing on chat thumbnails when a memory has multiple photos. Badge now shows the total count and is properly layered above the thumbnail glow effect.

Fixed the date of passing field being wider than the cause field in persona settings by overriding browser default minimum width on date inputs.

Stripped the browser's built in date picker chrome that was forcing the date field wider than the cause field

Switched the date and cause fields to use the same field styling as nicknames and profile fields, ensuring consistent widths

Fixed the date input being wider than the cause field by forcing it to match the same full width layout

Stacked the date and cause fields vertically instead of side by side to eliminate the persistent overlap issue

Fixed the date field border being clipped by letting it keep its natural width and pushing the cause field to its right

Fixed the date and cause fields overlapping in persona status by switching to layout layout, and removed a leftover empty media card from the Family tab

Fixed the date and cause fields overlapping in the persona status section by constraining inputs within their flex containers

Fixed the date input border being cut off in the persona status section by removing overflow hidden and slightly widening the date field.

Fixed date and cause fields overlapping in the persona status section by constraining field widths and preventing overflow.

Fixed overlapping date and cause fields in the persona status section. Date field is now compact and sits side by side with the cause field.

Moved the persona's status section into its own tab in the per person settings panel.

Added alive/passed away toggle with death date and cause to the per person settings panel in administration panel.

Added persona birthdate field to setup wizard step 1.

Added automatic age tracking for memories.

Changed video display from a list layout to a thumbnail grid matching the photos style.

Fixed photo and video preview opening at full resolution by constraining it to a smaller popup modal.

Moved Media management to its own dedicated tab in the admin panel.

Replaced the Music Library card in the Family tab with a unified Media card that includes Photos, Videos, and Music sections.

Renamed Session History to Games & Activities.

All time of day logic now uses the timezone from your persona config instead of a hardcoded timezone.

Greetings now only use the deterministic fast path response for the first greeting of the day.

Greetings now match the time of day — 'Good morning' in the morning, 'How was school' in the afternoon, 'How was your day' in the evening, and 'Shouldn't you be sleeping?' late at night

Farewell messages now match the time of day — daytime goodbyes say 'See you later' and 'Have fun', while nighttime goodbyes say 'Sleep tight' and 'Sweet dreams'

Saying 'I like this song' or similar compliments now tags the song for that person. Playback shuffles person tagged songs more frequently than general library songs for a personalized mix

When someone asks to play music, playlists tagged specifically for that person are now played first automatically — so child saying 'play my playlist' plays her tagged playlist instead of the general library

Fixed a bug where asking about family members would accidentally trigger their playlist instead of answering the question

Added a 2 second pause at full black between the farewell fade out and the splash page fade in for a smoother, less abrupt transition

Made the farewell fade out and fade in transitions longer and smoother for a more cinematic goodbye experience

Re imported all 12 playlists from scratch with 281 unique songs at . Total library size dropped from to . All songs are loudness normalized and thumbnails are properly sized.

Shrunk album art thumbnails from 300x300 to 96x96 and strip embedded artwork from mp3 files during import.

Cleared the entire music library to start fresh. Changed audio normalization from to so songs take up less space while still sounding great.

Added a person selector for music uploads and playlist imports so you can choose who songs are uploaded for instead of always defaulting to everyone.

Removed overly broad music play patterns ('how about X', 'let's do X', 'give me X', 'go with X') that would incorrectly intercept normal conversation.

Music interactions are now context aware.

Questions like 'what music do you like?' now properly list playlists instead of falling through to the AI.

Indirect music requests like 'I wish I could listen to some' now trigger playback. When a playlist was mentioned in the recent conversation, that playlist plays instead of random music.

When you reply with just a playlist or song name (like 'deep disco' after asking what music is available), it now plays that playlist or song instead of just talking about it.

Person checkboxes in song edit and playlist modals now show all family members from the persona config instead of only people already tagged on songs. You will always see everyone available to select.

Removed the Upload For dropdown from Songs and Import tabs.

Fixed the person selector in the song edit modal not responding when unchecking the All checkbox. Individual person checkboxes now properly become active and toggleable.

When 'All' is checked in the song edit modal, individual person checkboxes appear grayed out and checked. Unchecking 'All' activates them so you can pick specific people.

Removed album field from the song edit modal. The persons selector now shows checkboxes for all available people instead of a free text comma field.

The song edit button now opens a proper centered overlay modal instead of inline fields at the bottom of the list. Shows album art thumbnail and filename alongside the editable fields.

Overhauled the music library card in administration panel.

Music conversations now feel natural instead of robotic.

Fixed and issue where asking to play a specific song would show the correct song name but play a random track instead.

You can now ask to play a specific song, artist, or playlist by name.

Fixed seek and volume slider dots not being centered on their track lines — seek dot was slightly below and volume dot slightly above

Fixed VU meter peak indicator lines going outside the visible area — bars and peaks now stay fully within the visualizer bounds

Fixed VU meter peaks falling faster after skipping songs — animation loops were stacking up causing gravity to apply multiple times per frame

Increased VU meter sensitivity and responsiveness — bars now react faster to beats with snappier attack and tighter dB range

Completely redesigned the music VU meter to use proper ISO 1/3 octave frequency bands — each bar now represents a specific frequency range that matches what you hear, with much better beat responsiveness

Further reduced VU meter sensitivity so loud songs show proper dynamics instead of maxing out all bars

Reduced VU meter sensitivity so loud songs no longer max out all bars — the visualizer now shows proper dynamics across the full range

Fixed VU meter peaks going outside the visualizer bounds. New song downloads are now loudness normalized to 14 LUFS so all songs play at consistent volume

Fixed music player VU meter where the last 6 bars barely moved — capped frequency range at 16kHz and added stronger boost for higher frequencies so all bars respond visibly

Fixed music import failing on some songs by trying up to 3 YouTube search results instead of just the first one, which could be age restricted or unavailable

Fixed the music visualizer bars being lopsided — left bars were always maxed out while right ones barely moved.

Fixed a bug where songs with empty titles from Apple Shortcut export were being imported as broken entries.

Added search bars to both Songs and Playlists tabs in the music library.

The Import tab now automatically detects and resumes showing progress for and active playlist import when you reload the page or navigate back to it.

Added a new Import tab in the Music Library card that lets you upload .txt playlist files directly from the browser.

You can now tell persona to add the current song to your favorites or skip songs you don't like while music is playing.

Music now auto shuffles when it starts playing so you get a different song each time instead of always hearing the same one first.

The music player now scrolls long song titles that don't fit.

Long song titles now scroll smoothly instead of being cut off.

Reverted the music visualizer back to the version with 28 bars, smooth peak hold and gravity rain — the best looking iteration.

Bars no longer peg at full height — stronger compression and a 60% ceiling keep them dynamic and bouncy. Restored 28 bars for the fuller look.

Reduced equalizer to 16 bars, cut off dead high frequency bins, and added sqrt compression so bass bars don't always peg at max while mid and treble bars stay responsive.

Fixed music file upload failing with 'Can't find variable: axios' error by converting all music library HTTP calls to use the fetch API, which is the standard used throughout the admin panel.

Fixed music file upload not working on phones.

Fixed the music upload button not working on phones.

Fixed the music upload zone not responding to taps on phones. Uses a native label element instead of programmatic click, which works reliably on mobile browsers.

Added a music player that plays songs when you ask for music in chat.

Memories now support multiple photos.

When retelling a memory to the person it's about, the system now properly converts to second person with correct grammar (e.g. 'you were babbling' instead of 'you baby talking').

Fixed the gap between a reply with a photo/video thumbnail and the next message. The thumbnail now floats freely so subsequent chat lines wrap around it naturally instead of being pushed down.

Media thumbnails in chat now float to the right of the message text instead of pushing subsequent messages down.

The microphone now automatically mutes when you play a video in chat, preventing the system from picking up video audio as speech. It unmutes when the video is paused, ends, or the modal is closed.

The system now explicitly instructs the AI to never repeat memory text word for word. Memories are rephrased conversationally, like naturally reminiscing, instead of being echoed back as is.

Videos in chat now show a real frame from the video as a thumbnail instead of a generic play button. Thumbnails are auto extracted during upload using ffmpeg.

Fixed photo and video file inputs not clearing after saving a memory. The file picker fields now reset properly.

Fixed the date field overflowing outside the card by setting a fixed narrow width.

Made the date picker slightly narrower to prevent it from overlapping with adjacent fields.

Fixed the date field overflowing and appearing under the location field. All inputs now shrink equally to fit in a single row.

All four fields now sit in one row. Date uses its natural width while the others share remaining space equally.

Reorganized the Add Memory form fields into two clean rows: person and type on top, date and location below. Both rows use equal 50/50 split so fields align properly without gaps.

Removed the type hint text from the Add Memory form to close the gap between the field row and the affect section.

Made the date and location fields both auto width and grouped together at the end of the row, giving more space to the person and type dropdowns.

Reverted forced field sizing. Made the date column auto width and narrower so person, type, date, and location fit naturally in one row.

Fixed the Add Memory form fields so person, type, date, and location are all the same height and properly aligned in a single row without overlapping.

Fixed the Add Memory form layout so person, type, date, and location always display in a single aligned row using layout instead of flexbox, preventing fields from wrapping underneath each other.

Fixed the date field wrapping under the location field by making all four fields share the row equally instead of having a minimum width that caused overflow.

Fixed the Add Memory form so the person, type, date, and location fields align properly in a single row. Moved the type hint text below the row to prevent it from distorting field sizes.

Memories now support a location field.

Video uploads now auto extract the recording date from MP4 metadata, just like photos. When both photo and video are attached, the earlier date wins since that is when the memory happened

When uploading a photo to a memory form, the date field now auto populates from the photo's original date if available, so you no longer need to manually correct dates on older photos

Camera now self heals when a webcam fails or becomes unresponsive.

Fixed camera failing to start when the external webcam is unresponsive.

Fixed persona saying 'the girls' instead of 'you and child' when talking to spouse about shared memories.

Overhauled the keyword extraction system to correctly handle English word forms.

Fixed the keyword extractor generating nonsense words like 'treate' from 'treating'.

Fixed reactive statements like 'it was the best day of my life' or 'yeah it was great' getting 'I don't remember' responses during and ongoing conversation.

Added and auto detect button to the Add Memory form that analyzes the content text and pre fills emotional affect fields automatically, so you don't have to manually set feelings, intensity, and valence.

Fixed normal conversation like 'how are you' getting confused 'I don't remember' responses. Conversational messages now flow naturally without being blocked by memory lookup failures.

Parental alerts now catch natural phrases kids use like 'I met a boy named Dennis' or 'there is this girl at school'.

Fixed 'I missed you too' and similar past tense emotional messages getting confused memory responses instead of warm replies. Past tense forms like 'missed you' and 'loved you' now work correctly.

Memories about trips and experiences are now shared in a natural, conversational tone instead of being repeated word for word.

Fixed recipe and cooking requests getting 'I don't remember' responses.

Photos and videos now only appear alongside messages when the AI actually used that memory to build its response.

All story requests now tell a complete story from start to finish instead of the interactive turn by turn mode.

Fixed and issue where garbled speech to text input containing words like 'kill' or 'death' would trigger moral guidance mode, causing the AI to give irrelevant motivational speeches.

Fixed bedtime stories not working when a previous game session was still active.

Added three improvements: When speaking out loud, the microphone no longer picks up and transcribes the voice output VAD is suppressed during playback.

Fixed the Test Voice button to work in both browser and the Electron desktop app by trying HTML5 Audio first with AudioContext fallback

Fixed the Test Voice button not playing audio in the Electron desktop app by using AudioContext instead of HTML5 Audio element

Added a Test Voice button in the administration panel voice settings card that speaks a short test phrase through the speakers to verify the voice output is working

Switched the speech recognition engine from Whisper to NVIDIA Riva for better accuracy.

When the AI model loads with a smaller context window than expected, the system now detects this immediately and reloads with the correct size in the background.

Sensitive topics like death, drugs, and explicit content now always reach the language model for age appropriate handling instead of being short circuited as unanswered memory queries.

Fixed farewell responses not being captured in tests, improved memory relevance scoring so event memories ranked behind background context are correctly promoted, and added guidance detection for indirect bullying language like 'make someone cry' so children receive appropriate redirection instead of a blank response.

Added relationship aware privacy access control.

Implemented age based gradual content unlock for children.

Improved the privacy filter to use context aware phrases instead of single ambiguous words.

Fixed the privacy filter incorrectly tagging car/traffic accident memories as embarrassing.

Fixed a bug where the children were never detected as children.

Fixed the brain status flicker prevention to actually work on the landing page.

Fixed the brain/thinking status ring randomly spinning on the landing page even when the model is loaded and warm.

Fixed a response template that said 'I'm not sure I know who that is' when asked about trips or events, which didn't make sense. Now says 'I'm not sure I remember that' instead.

Added detailed documentation on relationship-based privacy controls and age-based content unlock.

Questions like 'What is the capital of France' and 'What's the tallest mountain' now correctly get answered as general knowledge instead of searching personal memories.

Fixed a crash when asking follow up questions like 'What about that time in Vegas'.

Fixed three issues: questions about general topics like electric cars no longer pull in unrelated personal memories.

When the system doesn't have a memory about something you asked, it now picks from 16 different natural sounding responses instead of repeating the same few phrases.

Activity invitations like 'can we bake cookies' or 'let's build a fort' now get warm nostalgic responses instead of 'I don't remember'.

Added a permanent reminder to every conversation that the persona has passed away, instead of trying to patch individual code paths.

Fixed a phantom memory that was hallucinated by the AI and incorrectly stored as a real memory.

Persona no longer suggests doing activities together like 'let's go out' or 'should we try that again' — since he's no longer alive, he can only share memories of when they did those things before.

Improved the 'no memories' case to explicitly tell the system not to invent answers, while still allowing general knowledge questions like 'What is the capital of France?' to be answered normally.

Fixed the system making up fake details when asked about things it doesn't have memories for. Instead of inventing answers, it now honestly says it doesn't remember.

Added rules to prevent the system from inventing memories, jobs, or details that don't exist in stored memories.

Fixed sibling questions like 'do you have a sister?' returning a wrong 'no' answer.

Renamed debug panel labels from generic 'v2' prefixes to descriptive categories like [session], [memory], [mode], [route], [safety], making it easier to follow what the system is doing during a conversation.

Fixed and issue where the system recovery process could load the main AI model with a smaller context window than needed, which would show a thinking spinner on the landing page until the model was re warmed with the correct size.

Fixed the page constantly refreshing in dev mode by configuring Vite's HMR WebSocket to use the correct external port instead of the internal container port

Made the on screen keyboard significantly smaller with reduced key padding, font size, and gaps.

Reduced the global font size across the entire app.

Debug panel now shows all debug messages from a chat response instead of only the last one. Previously each new debug line replaced the previous one, so you could only see the final debug entry.

Number input up/down arrows now sit right next to the input field as a seamless extension, instead of floating inside it.

Made the per person settings modal wider and more spacious.

Fixed the passkey input field height it now uses a fixed 28px height to ensure it renders compact.

Reduced the height of the passkey input field in the settings modal by shrinking vertical padding and font size.

Made the passkey input field less tall while keeping the same width.

Made the passkey input field shorter and more compact in the settings modal.

Made the settings modal more compact so it fits on screen without a scrollbar. Reduced spacing in mode selector, maintenance section, microphone options, and passkey area.

Fixed the milestones tab where the scroll arrows were breaking the page layout cards would disappear and content would jump when scrolling.

Added up/down scroll arrows to the milestones list, matching the same pattern used in the memories timeline for easier navigation

Removed the forced scrollbar and overflow clipping from the Add Memory card — it now displays at its natural height with no scrollbar or cutoff

The Add Memory area no longer stretches to fill the full page — it fits snugly around the form content

The Add Memory card no longer stretches to fill the full page height — it now sizes to its content, making all the spacing reductions visible

Significantly reduced padding, margins, font sizes, and textarea height inside the Add Memory card to fit without scrolling

Reduced the Add Memory page title and subtitle size to reclaim vertical space and eliminate the scrollbar

Reduced field margins, tab spacing, and action button margins on the Add Memory form to eliminate the scrollbar

Tightened the outer padding and header spacing on the Memories page to reduce overall height and eliminate the scrollbar

Tightened padding on the Add Memory form title and body to eliminate the scrollbar on desktop

The Add Memory form now uses a two column layout on desktop — core fields on the left, affect/feelings section on the right — so you can see everything without scrolling.

Removed the Children, Partner, Parents and Siblings category filters from the milestones dropdown since milestones are already assigned to individual people — you can filter by each person directly

Fixed the Children, Partner, Parents and Siblings filters in the Future Milestones dropdown so they actually show the right milestones instead of showing nothing when selected

Updated setup wizard to accurately reflect 12 steps (5 core + 7 optional).

Corrected voice prompt card count to 140 per set across all pages.

Added subject switch detection in the topic mismatch logic.

Added addressee detection that identifies when a query mentions a different family member than the active person.

Asking questions like 'What can you tell me about me?' now draws from rich profile data including personality traits, relationship dynamics, communication style, what I love about them, and inside jokes.

Previously, short memories would be dumped verbatim as the start of every response.

When asking about child or child being born, the response no longer starts with persona's own birth details.

When a memory carries emotional context like excitement or joy, responses now express those feelings naturally instead of just restating the stored text.

Math and coding mode no longer trigger on everyday words like times, number, homework, none, argument, sorted.

Emotional memories now carry structured feeling data so responses are more emotionally aware

Split memories page into three tabs: Add Memory, Recorded Memories, and Future Milestones

Fixed words running together in chat history

Photo thumbnails now overlay messages smoothly without pushing text around

Bedtime stories no longer appear in session history

Fixed session history showing unrelated conversations in bedtime story details

Removed blank duplicate entries from session history

Setup wizard now syncs with admin panel — entries persist across both

Setup wizard now shows existing personality data instead of appearing empty

Conversational reactions like "yes, it was a beautiful day" now get natural responses instead of pulling in unrelated memories

Fixed follow-up messages being misclassified, which prevented photos from loading

Fixed technical metadata leaking into conversation responses

Fixed birth questions pulling in other family members' details

Photos no longer repeat on follow-up messages

Built an automatic patch system that generates offline software updates and emails them directly to customers.

Added a version management dashboard for tracking software baselines and generating update patches.

Added sentiment detection for more natural follow-up responses

Background memories now stay in context instead of being used as direct answers

Fixed follow-up questions mixing memories from different people, and corrected grammar ("You were born" instead of "You was born")

Fixed "when were you born" showing children's details instead of own birthday

Updated development guidelines for better changelog descriptions

Fixed memory relevance scoring dropping to zero

Improved grammar in "I don't know" responses

Improved handling when no matching memories are found

Natural responses for unknown topics instead of making things up

Prevented making up answers when no memories match

Improved memory matching accuracy

Fixed possessive name handling in memory retrieval

Improved "what about" follow-up question handling

System now asks for clarification instead of guessing

Fixed topic matching for multi-memory responses

Added clarification questions for ambiguous topics

Fixed date handling in follow-up questions

Fixed photo thumbnails and follow-up memory retrieval

Improved filtering for recall-type messages

Photo thumbnails no longer push text around

Responses now naturally say "you" and "we" instead of using names

Responses now start more naturally

Fixed persona identification and best-match photo selection

Fixed birthday information accuracy

Fixed persona configuration loading

Fixed persona filter for family member queries

Fixed personal questions pulling in family members' details

Improved memory filtering to prevent irrelevant memories contaminating responses

Responses now directly reference the retrieved memory

Fixed session ending and improved synonym matching

Fixed pronoun resolution ("that", "this", "it") in follow-up questions

Fixed claude md persona model AI model to AI model instruct

Optimized AI configuration file

Fixed synonym matching and topic relevance checking

Added synonym matching (e.g., "wedding" finds "married")

AI responses now start naturally without visible anchors

Fixed photo matching and thumbnail display

Improved photo matching for short responses

Improved photo matching accuracy

Simplified photo matching logic

Fixed verbatim echoing and photo grid layout

Fixed topic switching and photo text overlap

Fixed "I just remembered" being classified as new information

Fixed memory response fallback and photo overlay

Photos now appear when asking about pictures

Status indicator now reflects AI readiness

Improved save script conflict handling

Unified character consistency detection

Improved pronoun handling for all family members

Fixed pronoun handling and photo display

Development sessions now persist for 24 hours

Fixed development mode access

Port dev gate to persona agi with cython and route

Landing page now requires facial recognition

Fixed AI model naming

Added loading screen while AI models start up

Patches now auto-fill their description from the git changelog, making update notes accurate and effortless.

Added cumulative patch generation — customers who missed updates can receive a single patch that includes everything.

Added safety lock to prevent accidental factory reset

Added development mode with person selector

Full swy code copy with persona data and 7 service patches

Added system synchronization

Unified speech-to-text, voice support, and security

Upgraded safety alerts and memory occasions

Upgraded farewell handling and context management

Upgraded to config-driven persona management

Added persona configuration system

Voice ID now loads birthdates from configuration

Face identification now saves snapshots

Parental alerts now appear in adult chat context

Elevated bullying and school threat alerts to high priority

Added toggles in support settings to disable IMAP polling and outbound email independently.

Fixed game mode indicator badge

Improved bedtime story paragraph flow

Fixed incomplete sentence detection

Improved sentence completion detection

Bedtime stories now use real names with proper paragraphs

Fixed bedtime story crash

Improved bedtime story sentence flow

Bedtime stories now use gender-aware language

Fixed bedtime story opening

Added editable canned responses in the support system for faster, consistent replies.

Support tickets now show the customer's tier and support level for context during conversations.

Added Pushover push notifications for new inbound support tickets.

Added bulk ticket management — select, delete, and mark tickets as spam in batches.

Launched an email-based support ticketing system with IMAP polling and file attachment support.

Updated Kickstarter campaign with video section, add-ons, stretch goals, early bird pricing, and social sharing.

Accuracy overhaul across website — fixed installer steps, add-on descriptions, language claims, and enrollment details.

Fixed names and capitalization in bedtime stories

Fixed GPU tooltip display

Added system status keepalive indicator

Added keepalive for faster AI responses

Fixed false positive in content safety filter

Fixed farewell detection for "thank you papa, good night"

Fixed game state persisting after farewell

Fixed bedtime story length for complete narratives

Improved bedtime story structure

Bedtime stories now tell complete narratives

Fixed "bed time" vs "bedtime" detection

Fixed goodnight story routing

Improved personal question classification (27 tests passing)

Expanded general knowledge topics and educational patterns

Fixed general knowledge routing for travel and family topics

Fixed sensitive state handling (61 tests passing)

Fixed ambiguous injury phrase handling

Improved grooming detection

Added comprehensive new parental alert triggers

Fixed classification of "I like", "I love", "I think" statements

Fixed bullying alert and information classification overlap

Fixed capitalization in message output

Fixed mid-sentence capitalization

Fixed capitalization after line breaks

Fixed mid-sentence capitalization

Fixed nickname handling in conversation priming

Fixed content framing for young children

Improved child-appropriate answers beyond just filtering

Fixed age-appropriate language enforcement

Toddlers now get magical, age-appropriate answers

Toddler answers avoid complex science explanations

Strengthened age-appropriate language for young children

Fixed general knowledge follow-up handling

Fixed greetings followed by questions

Diversified greetings and strengthened age-appropriate language

Rewrote age adaptation with concrete principles per developmental stage

Fixed chat history clearing

Fixed error notification frequency

Upgraded main AI model

Fixed "I don't remember" for general knowledge and sibling handling

Fixed follow-up reactions pulling irrelevant memories

Fixed face detection conflict with person selector

Fixed sibling memory handling and speech recognition restart

Fixed how are you memory query and wrong nickname for child

Fixed session ending on "thank you" and theme reset

Fixed audio runner riva streaming stt engine vad config child face folder

Fixed service conflict on rebuild

Added typewriter text effect (letter by letter)

Refined typewriter effect with natural pauses

Adjusted typewriter speed

Fixed typewriter streaming and cursor display

Memory pipeline audit fixes (round 2)

Second round of memory pipeline improvements

Fixed grace period and test failures

Fixed shared memories and math answer handling

Memory pipeline quality improvements

Fixed "what about" questions for general knowledge

Fixed wisdom memories bleeding into unrelated responses

Fixed "what about" subject extraction

Admin panel indicators now show disabled when services are off

Improved admin panel feedback messages

Added per-person breakdown for history deletion

Memory classification stats now persist across restarts

Fixed GPU profile for RTX 3090

Added model loaded indicator

Ported off-topic and personal preference detection

Wisdom memories now guide character rather than appearing directly

Fixed test cleanup and memory tracking

Fixed exact-match duplicate detection

Fixed query rewrite for user messages only

Fixed math detection and memory bleed in conversation

Optimized VRAM management on startup

Voice services now offload VRAM when disabled

Fixed context bleed — now always responds to the latest message

Fixed self-reference in photo gallery requests

Fixed memory tags leaking into gallery display

Fixed media gallery to block external content

Tightened anti-hallucination rules

Upgraded to 16K context window with 15 history turns

Fixed duplicate detection and sibling handling

Stopped unsolicited sibling mentions in responses

Fixed fragment handling, typos, and sibling tangent prevention

Improved conversation quality — no sibling tangents or typos

Fixed "play sport" not triggering game mode

Conversations now follow the thread naturally

Fixed admin panel modal display

Silent per-person reset without notification

Added session-based authentication to the support dashboard with IP lockout after 5 failed attempts.

Added per-person emergency reset with extended security

Reverted emergency reset from persona agi swy only

Redesigned Voice ID card with better upload and expand controls

Voice ID training now works even when system toggle is off

Added per-person emergency reset with long-press activation and 8-digit minimum passcode.

Fixed session tracking

Upgraded typo correction to 64K word dictionary

Enhanced typo correction and added database resilience

Broadened spouse parenting assertion add baby calm help keywords

More accurate chat timing metrics

Fixed spouse parenting assertion add persistent timing metrics

Added end-to-end performance benchmarks

Fixed typo correction mangling greetings

Added email address change feature for customers with automatic license token reissue.

Replacement tokens now track the reason (hardware change vs. email change) for better audit history.

All timestamps in the support dashboard now display in Mountain time for consistency.

Improved test cleanup

Fixed persona prioritize exclude remember from experience check

Improved memory retrieval by filtering noise words

Improved memory retrieval filtering

Reduced farewell response latency

Fixed typo correction for important words

Fixed typo correction preserving safety-critical words

Protected safety keywords from typo correction

Fixed typo correction keyword protection

Fixed test failures in safety and personality systems

Added typo correction for children's messages

Fixed chat history deletion

Cleaned up session management and admin panel

Eliminated hallucination when recalling memories

Fixed internal prompt data leaking into responses

Fixed memory priority and math/pet recall

Personal warm greetings with dynamic nicknames for kids vs adults

Greetings now respond instantly without AI processing

Comprehensive codebase audit: fixed filters, deduplication, and memory leaks

Added tier breakdown statistics and search to the customers dashboard.

Support dashboard installer now auto-detects port conflicts and prompts for alternatives.

Prepared voice training with natural sentences

Fixed persona memory retrieval model switching verification

Added comprehensive tone and feature test suite

Added realistic chat tests using real memories

Fixed general knowledge and memory detail accuracy

Fixed context mismatch — 6x faster first response

Added tier badges to the features page showing which features are available at each support level.

Added comprehensive 17-section legal disclaimer covering tier gating and arbitration.

Switched to verified instruction-tuned AI model

Added automatic GPU detection and adaptation

Improved natural memory formatting

Fixed pet recall with softer anti-hallucination and dynamic status

Switched to faster AI model — 3x faster with half the memory

Fixed pet memory handling and mode switching

Fixed AI restart and added timing checkpoints

Synced improvements: math detection, farewell handling, and emotional responses

Keyword relevance filter fix spouse hallucination model prewarm on mode exit clean persona search

Fast planning pass persona memory prioritization ttft 54pct faster

Fixed math mode persistence and character encoding

Added smart deduplication and age validation for memories

Improved filtering against manipulation and duplicate memories

Optimized speech recognition memory usage

Fixed debug output in chat stream

Reduced context window to fit in available memory

Upgraded AI model with structured memory reasoning

Added detailed help documentation per feature

Improved camera indicator badge positioning

Fixed blink detection badge position

Fixed admin panel status area layout

Fixed status indicator consistency and camera badges

Fixed dev bypass persona agi

Modals now stay open and degraded mode announces once

Redesigned admin panel with tab-based layout

Support dashboard mobile overhaul — cards instead of tables, responsive layouts, readable UUIDs on small screens.

Full dashboard overhaul with bundle history, hardware details, customer editing, and smarter token replacement.

Added descriptive help text across all support dashboard pages and sections.

Launched the customer support dashboard with git-backed data persistence and automatic syncing.

Swapped placeholder product image with a real screenshot of the system in action.

Added website analytics with country detection, session tracking, and a real-time dashboard.

Added real product screenshots throughout the website, replacing all placeholder illustrations.

Added personalization dashboard feature section to the website.

Added mobile-responsive navigation with hamburger menu.

Launched the Still With You marketing website.

Added scoring weights for all 20 memory types

Expanded memory types from 1 to 20 categories

Compressed existing child birthday 2 MOV to MP4 to

Added automatic video compression with orientation detection

5 response time optimizations: connection pooling, caching, and parallel search

Filtered transient messages from memory storage

Added settings font sm, settings font md, settings font lg variables All text now uses consistent sizing from these 3 tiers

Increased font sizes in Settings for better readability

When power draw is below 50% of limit , treat Software Power Cap as 'Power Saving' instead of throttling . This is normal idle power management, not real throttling.

Added context-aware disambiguation for coding keywords

Fixed mode loading timing and false positive detection

Added auto-refresh to hardware health monitoring

Fixed storage badge layout

Fixed SSD badge sizing

Added storage monitoring with warning thresholds (70% / 90%)

Admin panel now auto-loads saved layout

Fixed admin panel layout and camera text clipping

Improved hardware health labels

Added storage and USB monitoring to hardware health

Fixed admin panel font scaling minimum

Added all missing fonts across admin panel cards

Reverted admin panel to grid layout with unified fonts

Admin panel cards now fit their content

Improved admin panel card sizing

Fixed admin panel help card minimum height

Added missing fonts and improved help card height

Fixed admin panel component loading on mobile

Simplified admin panel card layout

Added font scaling support across all services

Unified admin panel font system with dynamic sizing

Fixed card overflow and added font scaling

Fixed font scaling for all elements including iOS

Added font size controls to admin panel header

Fixed admin panel stats layout

Added persona configuration with onboarding wizard

Kids under 12 see simple hardware warnings; older users see technical details

Added hardware monitoring with graceful degradation

Enhanced confidentiality warning messaging

Added comprehensive parental alert test suite (19 tests)

Added session-aware secrecy detection for child privacy

Added confidential handling for parental alerts with trust warnings

Expanded parental alerts: injury, peer pressure, sleep, academic, substance, and more

Added crush detection alerts and improved bullying detection

Fixed crush detection for all genders

Added scroll arrows to issues display

Error logs now include fix suggestions

Added milestone search, video filter, and statistics

Added life milestones with video support and event triggers

Added special occasion greetings (birthdays, holidays, etc.)

Moved wisdom seeds to conversation system

Added wisdom and values memories

Added life milestone messages system

Added age-adaptive communication (different style for toddlers vs teens vs adults)

Added confirmation dialog for switching modes

Added self-healing pattern learning for memory classification

Expanded pattern-based memory classification for relationships, preferences, facts, and events

Switched to smaller, faster AI model for memory classification

Added memory reset with cascading confirmations

Added core memory protection with admin tagging

Fixed memory ingestion in specialized modes with time-aware deduplication

Optimized memory processing during games and tutoring

Integrated staged memory retrieval pipeline

Extracted memory retrieval into staged pipeline

Added automatic error tracking and categorization

"What did we do?" questions now recall game sessions

Fixed session recall filtering

Split response handlers into modular packages

Improved admin issues display with detailed error messages

Added global error handling for memory and embedding systems

Migrated to modular dialogue architecture

Game mode now excludes dangerous topic requests

Removed deprecated chat modules

Fixed classifier for unified chat pipeline

Consolidated chat pipeline into unified architecture

Session recall now uses natural paraphrasing

Fixed session recall with mid-conversation support

Added "what did we talk about?" session recall

Game sessions now persist and end gracefully

Fixed feature icons and dynamic layout

Rebuilt memory display with dynamic layout

All admin panel cards now self-adjust to content

Fixed recall features and added themed confirmation

Added story and session recall settings to admin panel

Kids can now recall stories, games, and tutoring sessions

Fixed "Would You Rather" game reset

Expanded story and bedtime themes to 33 and 27 with child child tailored content

Added 176 riddles, 126 "Would You Rather" questions, and word chain game

Built extensive game library: 176 riddles, 126 WYR questions, 24 themes

Fixed games add persona like exit messages track riddles to avoid repeats add 30 riddles

Session audit log now shows full conversation history

Added session detail view with full information

Promoted Sessions and Issues cards in admin panel

Added self-healing watchdog and session history to admin panel

Added self-healing watchdog for audio service

Voice games now prompt to unmute speaker

Game exit detection is now more conversational and kid-friendly

Improved game mode detection and quit patterns

Added text games: 20 Questions, Word Chain, Story Building, Riddles, Would You Rather

Fixed age verification, output quality, and parent references

Help chat now explains features naturally instead of quoting documentation

Admin panel uses smooth scrolling on mobile

Improved help chat with dedicated math and coding documentation

Improved help chat button placement

Fixed help chat preventing identity information leaks

Help chat now provides detailed system overviews

Fixed help chat handling off-topic and identity questions

Improved help chat for natural language queries

Improved help chat accuracy by preventing fabrication

Smart mode continuation with AI-based context checking

Added off-topic detection: casual conversations now suggest switching from math mode

Injury, illness, and bullying topics now exit tutoring mode for emotional support

Tutoring mode now stays active unless clearly exited

Added advanced math terms for better mode detection

Fixed variable initialization error

Fixed system understanding: math and science are for everyone, not "grown-up things"

Improved audio handling: microphone priority, dynamic device switching, and resampling

Separated math and coding mode toggles in admin panel

Admin panel cards now auto-paginate based on screen size

Compacted card title spacing

Compacted admin panel card titles

Improved camera card layout

Admin panel now uses full words instead of abbreviations

Fixed model slider control size

Fixed slider controls in narrow cards

Added configurable echo cancellation and search tuning

Compacted model and voice cards for narrow displays

Fixed encryption card display

Redesigned admin panel cards with improved iOS support

Added tappable tooltips to all admin panel labels

Redesigned camera card layout

Added card dimension indicator in edit mode

Fixed camera card row stretching

Added snapshot settings with live preview

Fixed photo gallery year filter and search

Photo gallery now handles misspellings and family/year filters

Fixed photo gallery responses per family member

Added playful roleplay actions (rummages, scrolls, flips)

Added configurable IR thresholds and fixed dual-camera streaming

Launched facial recognition v2 with improved IR thresholds

Fixed IR camera pairing for primary camera

Added safety check for noise level when listening is disabled

Noise warnings now only show when audio is enabled

Fixed liveness detection configuration reload

Fixed IR camera pairing for primary camera

USB speed warning now only shows for external cameras

Added USB speed warning and single-shot camera mode

Added IR single-shot mode for USB 2.0 cameras

Improved "could you" and topic detection patterns

This reverts commit 638cf1e22e2f5eacf68e7c50b12e7606714be7c5.

Fixed photo gallery search patterns

Added roleplay action sanitizer for text cleanup

Fixed roleplay narration restrictions

Added photo gallery with natural language access ("show me photos")

Added paired IR anti-spoofing from same physical camera

Fixed comprehensive tests for greetings, mode exits, and farewells

Cleaned up duplicate code with centralized configuration

Added echo cancellation to prevent self-listening

Added voice synthesis toggle and fine-tuning recording script

Fixed GPU memory tracking accuracy

External camera now used as primary

Fixed camera selection for RGB and IR pairing

Fixed memory database connectivity check

Added memory database status indicator

Fixed database restart and IR settings persistence

Fixed voice synthesis error and added loading indicator

Fixed save script path variable

Added service health checks and responsive memory interface

Added persona dev auth prompt on facial detection grainier dot preview 48px

Fixed IR toggle and adjusted admin panel fonts

Enhanced security with rate limiting and obfuscation

Updated status labels and added voice enrollment indicators

Renamed Speaker to Voice ID for clarity

Renamed voice identification labels for clarity

Added blink liveness toggle and anti-spoof controls

Fixed anti-spoof priority: IR default with blink fallback

IR liveness prefers USB IR with blink detection fallback

Added IR duty cycling toggle with camera indicator

Added IR liveness status to admin panel

Added IR liveness detection with improved preview quality

Fixed race condition in person switching

Added encryption status indicator

Fixed memory export for large collections

Added memory export to backup script

Rebuilt emotion editor with proper sizing

Fixed advanced options layout

Improved advanced options layout

Fixed emotion confidence buttons overlap

Consistent font size across all filters

Fixed media dropdown text clipping

Improved task filtering and dropdown arrow visibility

Applied theme colors to network setup

Themed remote access modal and fixed URL overflow

Simplified Voice ID tooltip text

Fixed Voice ID toggle in admin panel

Mode-specific remote access prompts

Improved remote access dialog styling

Fixed remote access button and checkbox styling

Fixed network hostname resolution

Added network debugging tools

Improved remote access button styling

Themed header indicators and fixed mode persistence

Fixed remote access dialog behavior

Remote access info now auto-shows once per user

Added remote access for coding and math modes

Fixed info button and removed duplicate

Fixed remote view prompts for coding and math

Added remote companion feature with simple access URLs

Added code export to shared folder

Added code block download button

Fixed false positive: academic topics no longer flagged as "grown-up"

Reduced false positives in memory issue detection

Fixed false memory issue popup in coding mode

Tutoring sessions auto-save every 3 turns

Fixed theme flickering during mode changes

Fixed code block rendering

Added multi-line input with code block copy button

Optimized AI responsiveness in specialized modes

Speed optimizations for math and coding tutoring

Fixed AI model switching efficiency

GPU tooltip now shows all running AI processes

Increased context and history for faster tutoring responses

Fixed essential AI model staying loaded for 24 hours

Fixed specialized mode model management

Switched to friendly model labels: Conversation, Memory, Math Tutor, Code Tutor

Unified model names across the interface

Renamed to Math Tutor and Code Tutor

Simplified model configuration card

GPU tooltip shows loaded models with memory usage

Simplified model display and GPU tooltip

Moved model display to GPU tooltip

AI models stay warm; specialized models load on demand

Added GPU monitoring for accurate memory tracking

Updated encryption card layout

Replaced version card with header button

Added encryption management card to admin panel

Security card now shows protected configuration count

Added encryption management to admin panel

Added memory encryption toggle to admin panel

Updated tests for math and coding mode detection

Fixed encryption across all memory operations

Added encrypted storage for memories and chat history

Added development history to backup

Updated development tools installer

Fixed deployment paths and configuration

Voice enrollment now supports m4a, mp3, flac, ogg, and aac formats

Voice ID retraining now runs in the background

Fixed per-person voice sample limits and auto-retraining

Cleaned up Voice ID configuration

Cleaned up facial recognition code

Fixed status indicators on startup

Added AI fallback for future-proof message classification

Fixed math mode: "problems with friends" no longer triggers math

Added natural language detection for math and coding modes

Fixed specialized mode initialization errors

Fixed personality profile formatting

Consolidated duplicate age calculation functions

Verified memory ingestion after cleanup

Cleaned up invalid memories and verified improvements

Added detailed tooltips to person settings with 10-minute timeout

Merged Features and Speed into single card

Added flexible grid layouts for feature cards

Fixed missing cards in saved layouts

Merged Tools and Logging into Debug card

Preserved user customizations when removing cards

Updated admin panel layout version

Added warning for large context windows

Merged Search into Context Memory card

Removed duplicate memory card

Fixed word stuttering in responses

Fixed AI streaming duplicate text

Added fun loading messages when switching AI models

Connected specialized math and coding AI models

Upgraded voice synthesis and expanded context window to 24K

Added 16K context window with configurable memory retrieval

Fixed tooltip positioning at different zoom levels

Fixed tooltips on mobile touch devices

Fixed tooltips escaping card boundaries

Fixed tooltip alignment

Fixed tooltip positioning and status alignment

Added fine-grained card resize controls

Layout export now saves to server

Fixed history card sizing

Preserved saved layouts and added export button

Redesigned admin panel layout with voice training button

Fixed stats card minimum height

Added edit lock for admin panel layout customization

Admin panel cards now draggable and resizable

Redesigned admin panel security layout

Rebuilt admin panel with 3-column layout and quick toggles

Rebuilt admin panel with 5-column weighted grid

Rearranged admin panel for better organization

Fixed version history display for quoted text

Version history now shows all entries

Version history entries now show full details on click

Fixed version history data source

Upgraded version history with API and cache fallback

Increased person settings dialog height

Unified personality fields across all people

Compacted person settings and auto-open on login

Added clear history per family member

Redesigned person settings with dynamic people list

Added person settings with nicknames and personality profiles

Consolidated nickname management

Added personality profiles that shape conversation style per person

Fixed configuration and added error logging

Moved voice recordings to persona agi and create cleanup script

Removed legacy code, consolidated chat tracking

Simplified character break recovery

Added 15 fallback responses for character breaks

Added fallback when AI breaks character

Fixed false camera freeze detection

Added Voice ID system toggle with status indicators

Fixed memory deletion with error feedback

Fixed tooltip going off screen

Cleaned up status indicator styling

Separated admin panel cards for better organization

Fixed status indicator hover effects

Fixed status indicators after authentication

Voice ID tooltip now shows enrolled names

Voice ID model now loads on startup

Added Voice ID status indicator

Added automatic folder setup

Simplify enroll command . enroll child

Created voice sample folders for all family members

Voice ID automatically graduates infants at age 3

Voice ID requires face+voice match; added infant mode

Added voice identification with folder-based enrollment

Updated persona interface to wait for graphical session

Rename kiosk service to persona interface

Fixed frontend URL configuration

Added auto-start with display fallback

Chat blurs when a different person is detected for privacy

Removed "under development" text

Cleaned up keyboard, modal, and security transitions

Specialized modes persist with idle timeout and grace period

Strengthened mode continuation detection

Fixed import error in admin routes

Fixed offline warning badge color

Added restart button for offline services

Updated performance thresholds from benchmark results

Improved service error messages with specific reasons

Added real-time clock to admin panel header

Fixed admin panel tooltip positioning

Fixed settings colors and tooltip clipping

Fixed chat history leaking across sessions

Fixed C language detection for coding mode

Added timing metrics for specialized modes

Fixed mode theming and bash detection with response fallback

Backend improvements: safety, validation, and response handling

Frontend refactoring: state management, polling, and error handling

Improved admin panel styling

Added GPU estimation when monitoring unavailable

Added per-service GPU memory breakdown

Added full GPU monitoring and character consistency tracking

Added service status indicators with tap-for-details

Added detailed statistics per person

Added timing statistics charts

Fixed statistics layout and tooltips

Fixed face detection status check