When the assistant's instructions ran to a thousand lines
Every time you talk to XNeuronal, the model first rereads its instructions. On the morning of 21 May they ran to 1,043 lines: rules, counter-examples, special cases added one at a time over months, and every addition knocked another one over. That day, everything was rewritten as 408 lines that describe what the app is, rather than what to say in each situation. Every tool now carries its own notes, read only at the moment it is used. And the assistant learned to say hello once a day, not with every sentence. The result: steadier replies, fewer empty "noted", and a tone that holds.
One "noted" too many
Harriet is telling the app about her week on the way home. "My brother Callum is coming on Sunday, he's given up sugar, I need to find him a dessert." The app answers "OK, noted." Two sentences later, Harriet asks for a dessert idea, and the reply suggests a treacle tart, without a word about Callum. The fact had been kept. It had also been forgotten, within the same minute.
Scenes like this were not rare in the spring. One sentence could hold three memories and leave none behind, as we have told. But there was a duller flaw, harder to photograph: the assistant did not have the same character from one turn to the next. One day it commented on every piece of information with a "noted", the next day it stayed quiet. Sometimes it recapped what you had just said, sometimes it asked the right question. And in the middle of a conversation, it could lose track of what had been said two sentences earlier.
We first looked for the cause in the memory. It was somewhere else: in the text the model rereads before every reply.
A thousand-line manual
A language model knows nothing about XNeuronal when it arrives. Everything it knows about the app, its memory, its tools and the way it should speak lives in a text it is handed at the top of every exchange: the instructions.
Those instructions had grown the way house rules grow: by addition. A bad reply in testing, one more rule. A case handled badly in real use, one more example, with the wrong answer crossed out and the right one next to it. On the morning of 21 May the text ran to 1,043 lines. It held thirty-two lines of examples of the "don't say this, say that" kind. It held the rule that separates what is kept silently from what is confirmed out loud, repeated in five different sections, each time slightly differently. It held 175 lines describing, field by field, the structure of a card, with an example for each kind of memo: recipe, shopping list, reading note. And it held fifty lines on how to order an illustration image, when that command had been taken away from the assistant nine days earlier: since 12 May, the server picks the image on its own. Nobody had removed the manual.
Nine days earlier, too, a map of the memory had been placed at the top of that text to tell the model where each piece of information goes: useful, and a few dozen more lines.
The real problem was not the length as such. It was that each special case contradicted another. "Always confirm when asked to remember" sat next to "never say noted". Every rule had been right in its context, and the model received all the contexts at once. It applied some of them, not always the same ones.
Why a longer manual makes the assistant less reliable
A language model does not read its instructions once and for all. It rereads them in full on every turn, together with the conversation history, the active memory and the description of its tools. All of that forms a single block of text, and that block weighed on the order of fifty thousand tokens per turn, a good half of it for the instructions alone. A model's attention is not uniform over a block that size. What sits in the middle counts less than what sits at the beginning and the end. And the thread of the conversation, what you said two sentences ago, sat precisely in the middle, wedged between a thousand lines of rules and the list of tools.
That is what explains Callum's dessert. The fact had indeed been written to memory. But on the next turn, in the mass being reread, it weighed less than a sample couscous recipe. The model had not forgotten: it had not looked there. And every turn took longer to process, which every future addition would have made worse.
Saying what is true rather than what to do
Late in the afternoon of 21 May, the text was rewritten from scratch. Not shortened: rewritten, on a different principle.
The old text said what to do: in this situation, reply this, call that tool, avoid that phrase. The new one says what is true. It answers thirteen questions, in order: who you are; what XNeuronal is and is not; why it exists; how its memory is built; how a memory is classified; how its content is written; which tools exist; how they chain together; when to keep something silently and when to confirm out loud; why nothing is ever erased; how to speak; the few situations that deserve separate treatment; and what the server adds on every turn.
The bet fits in one sentence: a model that has the world explained to it works out the right behaviour better than a model that has the answers dictated. Take memory. The old text listed filing rules. The new one explains that the app's memory imitates three systems of human memory: what we know, the lasting facts about your life; what we have lived, the events set in time; what we have to do, the actions to trigger later. It says that the first category is silent, that no card opens when you mention your sister's name, and that the third is always visible, because you expect to see the reminder set. From those three sentences follow dozens of cases the old text listed one by one.
The rule of silence, the one that was repeated five times, is now written only once, in two paragraphs. You tell it about your life: the app keeps it without comment and reacts to the content, asks the useful question, points out a contradiction. You explicitly ask for an act of memory, "note this", "remind me at four": it confirms in one short sentence. It is the same rule as before. It no longer has a rival.
The result is 408 lines. Sixty-one per cent fewer. The block reread on every turn was roughly halved, and the conversation history got its room back.
Every tool carries its own notes
Part of what disappeared from the instructions was not deleted. It moved.
The assistant acts on the app through tools: create a memory, edit one, search for some, evolve one, set a reminder, search the web, show a card, declare that someone else is speaking, record a behaviour rule. Nine tools, that day. Each is described to the model with its parameters, and that description is read at the moment it is about to use it.
That is where the 175 lines of card structure went, along with the one-to-ten importance grid and the table of the four reasons a memory can evolve: a fact that changes, a fact that was wrong, a problem resolved, a piece of information you ask it to forget. Fifty-four lines added to the tool notes, six hundred and thirty-five removed from the instructions. The same information, but consulted only when it is needed. The assistant does not need to know the structure of a recipe card to answer "hello". It needs it when it creates the card, and that is when it is shown.
The complete recipe example, the one that groups ingredients by family, meat, vegetables, spices, rather than dumping them in a heap, is still there, in the notes for the field that describes the card. Read when you ask for a recipe, not when you ask the time.
What was thrown out, and what we refused to keep
The block about images was deleted without replacement: it described a tool the assistant no longer had, since the server builds a card's illustration on its own from the title, the teaser and the chips. We could have kept it "just in case". It would have gone on being reread on every turn for nothing.
Three other options were set aside. Keep the text as it was: the symptom persisted and every addition made it worse. Cut it by a third while keeping the "don't say this, say that" style: the redundancies would have stayed, and with them a way of treating the model as a parrot to be trained. Keep the card structure in the instructions: that meant making it read an assembly manual before every "hello".
The decision is reversible, and that is written down in black and white in the project's decision log. Two commits to undo, no database touched, no interface changed. If the new text turned out to be worse, rolling back takes a minute.
Hello once a day
The same day, a little earlier, the assistant had learned something smaller and more visible: to say hello.
Until then, it could skip the greeting entirely when you opened the day with a direct request, or on the contrary answer "Good morning Harriet" on every exchange, as if it were meeting you for the first time every ten minutes. Neither is what someone who knows you does.
The server now knows whether the conversation being opened is the first of your day: it looks at the last archived conversation and checks whether it dates from today, on Paris time. That information is added to the context of every turn, along with the time. If it is the first conversation of the day, the assistant opens with a good morning before six in the evening and a good evening after, followed by your first name if it knows it. It takes the initiative even if you go straight to a request: "hi, tomorrow's weather" gets "Good morning Harriet, let me look." Later in the day, it does not start again. If you greet it, it simply greets back. If you get straight to the point, so does it.
Two details of this ritual say something about how the app is built. First: if it does not know your first name, it says a plain "good morning". Never "good morning dear friend", never a name guessed from memory. Second: when the answer is going to take a few seconds, the short phrase the assistant says right away, "OK, let me look", carries the greeting too that morning. The courtesy arrives in the immediate audio, not only in the full reply that follows. And if the phone has been lent, it is the name of the person speaking that is used, not the owner's.
In the old text, this ritual took up a whole section, three numbered rules and six examples. In the new one, three paragraphs, which do the same thing.
What changes when you talk to it
None of this shows in the interface. No button moved on 21 May. What changes is the steadiness.
You mention that your brother has given up sugar: the app no longer says "noted". It reacts to what you just said, or it stays quiet, and the fact is kept. You ask for a dessert two sentences later: it has the thread in front of it, because the thread is no longer buried under a thousand lines. You set a reminder: "OK, reminder set for four", one sentence, not a recap. A card opens: one accompanying phrase, "there, it's ready", never silence over a screen that has just changed. And the conversation keeps the same tone from one day to the next, short, informal, without a presenter's enthusiasm, because that tone is described once rather than prescribed twenty times.
What the app had learned two days earlier about the ephemeral, showing a weather forecast without keeping it, found its place in the new text as a principle, in three lines, not as one more case. And the rules you give it yourself, "stop bringing up that subject", "always remind me of this", are described as what they are: a lasting instruction that you add to its own, recorded by a dedicated tool, reread on every turn in its active memory.
What we are watching
A declarative text makes a bet on the model reading it: that it will work out the right behaviour from a clear mental model. Not every model reasons equally well, and the one running in production that day is not the one the text was first written for.
The two places where a regression is expected are known. A sentence that holds several facts, one of which could slip through. And the difference between "remind me" and "remember that", which separates a reminder from a plain fact to keep. For the first, the second pass introduced on 18 May, the one that recounts the facts in a sentence after the reply, stays in place as a safety net. For both, the rule is set: if the model misses a behaviour, it gets a targeted instruction, one line at the exact spot where it is missing. We do not go back to the manual.
Harriet, for her part, will mostly notice one thing. On the Sunday Callum comes round, the app will suggest her a sugar-free dessert, and it will not have said "noted" to know it.