When one sentence holds three memories
Until 18 May, a sentence like "by the way, my wife's name is Claire, my daughter is eleven and my son eighteen" could leave no trace at all: the model answered warmly and filed nothing. That day changed the way the memory listens. Every piece of information gets a category, what you know, what you lived, what you have to do, and that category decides whether it is shown or kept in silence. A second pass, after the reply, counts the facts in a sentence and catches the ones that are missing. When it is time to remember, the memory searches from several angles at once, and what it finds warms up its neighbourhood. And the constellation gives each kind of memory its own shape.
An ordinary sentence, and nothing behind it
The neural memory is built for one precise use: you talk, and the app retains what deserves to be retained, without being dictated to. A week earlier, search had learned to follow the links between memories. What remained was the start of the chain: what comes in.
The flaw was observed in the database, not guessed. Someone had opened a conversation with "hi, listen, my wife's name is Claire, my daughter is eleven and my son eighteen". The model had answered kindly, asked a question about the children, and created no memory. Not one out of three: none. The sentence opened with a greeting, it contained no command, and it delivered three pieces of information at once. Each of those traits makes a language model hesitate to call its tools; together, they were enough for it to call nothing.
That kind of flaw is not visible. A reminder that fails to get set is noticed at once: nothing appears on the screen. A fact that fails to get retained is discovered three weeks later, when you ask "what's my wife's name again, for the booking?" and the answer is a question. And the app had chosen from the start not to say "noted" after every piece of information slipped into a conversation. The silence was deliberate; it made the flaw invisible.
Late afternoon: remind me, remember that
The first change is a matter of grammar. "Remind me on Thursday to pay the rent" and "remember that the gate code is 4820" differ by a word or two, and those words change everything. The first asks for an action at a future time: a reminder, set in the calendar, with a notification and a brief spoken acknowledgement. The second asks for nothing at a future time. It asks to know. It is pure memorisation, and it produces a fact, not a reminder.
Until then, the instructions treated both the same way, and the outcome depended on the model's mood: sometimes a reminder with no date, which is useless, sometimes a "noted" with nothing behind it. One question now settles it: is the person expecting something at a given moment? Yes, it is a reminder. No, it is a fact.
And a fact is retained in silence. The model does not say "noted", it reacts to what it has just been told, like someone who listens. The same rule applies to sensitive information, a code, a login: retained without asking for confirmation and without comment. It is the person's decision to entrust it, not the app's place to comment on it.
Six o'clock: counting the facts in a sentence
The heart of the day is the three-fact sentence. Two answers were given, one in the instructions, one in the code, because the first was not enough.
On the instructions side, the rule became explicit: one at a time. When a sentence contains several facts or several people, the model creates as many memories as there are distinct elements. It never merges them, it never leaves one out. A greeting excuses nothing: it may answer warmly, it must still retain. And before answering, one question: how many distinct facts in what I just heard?
That instruction helps. It guarantees nothing. A language model follows an instruction with a probability, not a certainty, and for a piece of information you will never hear twice, a probability is not enough. A safeguard in the code was needed.
That safeguard is a second pass. Once the main model has answered and its tools have run, a checker receives two things: the user's sentence, as is, and the list of what was already retained during that turn. It has one task: spot the facts and people the sentence contained that the list does not cover, and add them. It has a single tool, the one that creates a memory, restricted to two types: a fact or a person. It never creates a reminder, an event, a memo or a task. Those things are visible and remain the business of the first pass. The checker corrects only the silent memory.
It has other prohibitions. It does not speak: if it has nothing to add, it returns nothing. It never makes a card appear. It does not duplicate: the list of what already exists is there for that. And it creates only when certain: a first name that could be a nickname, and it abstains. It invents nothing that is not in the sentence.
What it catches follows the same path as everything else and joins the list of the turn's creations. If it fails, the turn does not fail: the error goes to the logs, the person never hears of it. The price is one more call per turn of speech, and a little waiting. It is the price of a memory that does not let a sentence slip by.
Seven o'clock: three ways of retaining
The next step put words on a distinction the app had been practising without naming. Work on human memory has long described distinct systems for what you know, what you lived and what you have to do. The model's instructions now begin there.
There is what you know: durable facts about the person and their world, a first name, a job, an allergy, a code. This is the knowledge the model will draw on in the next exchanges, and it is always retained in silence. Never a card, never a "noted".
There is what you lived: events and memories situated in time. Their visibility depends on the case. A dated event is shown, because the person is planning and needs to see. A memory slipped in passing, "yesterday we did this", is not shown, unless the person clearly wants to look it up later.
And there is what you have to do: tasks and reminders, always visible. The person expects to see and to be notified.
The question to ask before every memory follows from this: this piece of information, do I retain it, or do I show it? By default, retain. The same distinction applies to the summary that appears when several things are created at once: only visible creations count. Two facts retained in silence during an ordinary conversation no longer bring up a card that nobody asked for.
Searching widely, and keeping the neighbourhood warm
The second half of that step concerns the opposite moment: when the person asks for something that depends on the memory. "What's my daughter's name?", "where did I put the passport?".
A living memory does not answer those questions by looking up a single line. A fragment of a memory lights up the whole: saying "my wife" brings back the first name, but also the birthday, the trip in March, the allergy. The instructions now ask the model to do the same. Instead of a single search, it launches two or three in parallel, from different angles: by the person mentioned, by the topic, by similarity of meaning. Then it crosses the results. "You told me you were in Rome with Claire in March" assembles an event, a person and the link between the two; the answer is the synthesis, not an isolated memory.
And what the search finds leaves a trace. After every search that finds at least one memory, a background job refreshes the last-mentioned date of the three best results and of their direct neighbours in the graph of links, at most two neighbours per result. Nine memories at most, deliberately few, so as not to skew the sorting done by the job that tidies the memory. The effect: the topic just discussed, and what surrounds it, stays in working memory for the following turns. This propagation runs in the background; the answer is not slowed down by it.
A shape for each kind of memory
The constellation, that map where every memory is a dot and every link a thread, received the same evening a visual distinction it had been missing. Until then, every dot was a circle.
Every dot keeps its halo and its glow, identical for all. But the core of the dot now has a shape that says its category. A solid circle for a fact. A circle ringed in white for a person. A diamond for an event. A hexagon for a memo, the shape of a note you consult. A triangle pointing up for a task. A small hollow diamond for a reminder, a one-off signal. And a star for topics, those memories of a separate kind born a week earlier, which tell nothing themselves and point to the others. A legend under the map recalls the code, and the three categories become readable at a glance: the circles are knowledge, the diamonds and hexagons what was lived, the triangles and hollow diamonds what remains to be done.
Past seven: the thumbnail that betrayed the silence
The last change of the day came from a voice test, and it is small. Under every reply from the assistant, when something was created during the turn, a thumbnail offers to open the card. But the thumbnail also appeared when the model had silently retained a fact or a person over the course of an ordinary conversation. The conversation was ordinary, the reply was ordinary, and an "Open the card" thumbnail said the opposite: something was filed, come and look. It was exactly the "noted" we had forbidden ourselves to say, written on the screen.
So at the end of the turn, the server re-reads the nature of what was created, and passes on to the phone only the visible memories. A fact retained in silence stays silent all the way.
The same test had revealed another habit. When the three family members are known and the person says "we went to the seaside as a family last week", the model tended to copy into the memory of the weekend everything the people's cards already said: the wife's first name, the daughter's age, the son's age. The instructions now lay down a rule: the text of an event describes the event, where, when, what happened; the people are referenced separately, by first name, and that field is what will bring back their cards when it is time to remember. The test to apply: if this memory disappeared and I asked the person again, what would they tell me that is new? That alone goes in the text. The rest comes from the crossing.
What the day taught
18 May did not make the memory bigger. It made it more faithful to the way we listen. You do not retain a first name the way you retain an appointment. A sentence can hold three things to keep, and they have to be counted. A partial cue should light up a neighbourhood, not a line. And what is retained in silence should stay that way, down to the small signals of the interface.
The checker is perhaps the most revealing piece of that day. It turns an instruction, which the model follows most of the time, into a guarantee, which the system holds always. For a piece of information you will hear only once, the difference between most of the time and always is the whole difference.
The next day, another day would tackle the thread of the conversation in progress, that short-term memory lost every time the screen went dark. But since 18 May, "hi, my wife's name is Claire, my daughter is eleven and my son eighteen" leaves three memories, in silence, and "what's my wife's name?" gets an answer, with what surrounds it.