My Collaborator Has No Memory Except the Notes I Keep for It

SCIENCE & TECHNOLOGY · SEPTEMBER 11, 2026

The programming partner I have worked with every day since June has no memory. I mean that literally, not as a complaint. It is an AI coding assistant — Claude Code, from Anthropic — and every time I open a new session it begins with a fresh context, knowing nothing about yesterday's work, the decisions we made, or the mistakes it caught. What it knows about my project, it knows because it reads it from a file at the start of each session. If a fact is not in that file, or in the code itself, or in the git history, then as far as my partner is concerned that fact does not exist.

I did not expect this to be the most interesting thing about the arrangement. I expected the interesting part to be the code. But three months in, the thing I would most want to tell someone starting the same way is not about code at all. It is about what happens when the only continuity in a working relationship is a document — and what that document does when nobody is looking.

How Does an AI Coding Assistant Remember Anything?

Three ways, in my setup. The first is a file in the project's root folder, named CLAUDE.md, which the tool reads in full at the start of every session. Anthropic's own documentation describes it as the place you write "what you'd otherwise re-explain," and recommends keeping it under 200 lines. Mine is 1,255 lines and about 210 kilobytes — down from 685 kilobytes at its peak in late August, when a routine check of my usage showed that every session was starting a good way into its budget before I had typed a word, because it was reading a small book first.

The second is a folder of notes the assistant keeps for itself: one fact per file, with a short index it reads at startup and the full notes it opens on demand. Mine has 231 of them. They are plain text, and I can read every one. About a third are corrections — things I told it once that it wrote down so it would not need telling twice. One of them, which I did not write and would not have thought to, records that on two separate occasions, when I had asked it to help me sort and label a batch of travel photographs, it had described me eating and enjoying a dish I had only photographed, and it had invented the sensory details. (One was Brussels sprouts. I don't like Brussels sprouts.) The note calls this "the standing failure mode, not a one-off," and every session since has read it before describing anything I did.

The third is the git history: every change, with a message saying why. Between the three of them, a new session can reconstruct a surprising amount. What it cannot do is remember. It can only read.

What Happens When the Notes Go Stale

Here is the thing I learned the expensive way. A note that is read completely, trusted completely, and never remembered from one reading to the next does not degrade the way a human's memory does. It does not fade. It stays exactly as confident as the day it was written, while the world it describes moves on.

The project has a section in that file called "Current Fleet State" — which of my trading strategies are live, which are simulated, how much each holds. For a while its heading said authoritative. On August 7 the assistant read that section, found a problem described there, and reported it to me as though it were happening now. It was not; only the note still said so. On August 13 the same section was still describing one strategy as ninety-nine percent of my brokerage account two days after I had sold every share it held. Nothing was wrong with the assistant on either day. It read what I had left it, and believed it, which is precisely what I had asked it to do.

The heading now reads: "The live account is the source of truth, not this section." Underneath it are three commands to run against the real brokerage account before making any claim about money. The word "authoritative" was removed, with a note explaining that it was the thing that made a stale copy look trustworthy. I keep coming back to that sentence, because it is true of a great many documents besides this one. The problem was never that the note was wrong. The problem was that it did not know it was wrong, and neither did its reader.

There are smaller versions of this all through the file, and the file itself now records them as lessons. A note that announced a branch had been deleted, written a day before it actually was — the deletion was in the plan, and the plan got written up as the deed. A strategy I retired that turned out to still own a piece of shared plumbing (a cache of dividend dates that only it ever refreshed), so that switching it off silently froze a number a real-money strategy was reading, with no error anywhere; the note that retired it did not know the plumbing existed. Each of these is now a paragraph in the file, in the assistant's own words, addressed to its future self.

The Answer Was Less Prose and More Machinery

The instinct, when the notes go stale, is to write more notes. That is what we did for two months, and it is how a file gets to 685 kilobytes. The thing that actually worked was the opposite: take the facts that can go stale out of prose and put them somewhere that checks itself.

Three examples. The fleet-state section stopped saying what is live and started saying how to find out. A rule that every change to a strategy must also update its dashboard page — which I had asked for, and which prose alone kept failing to deliver — became a small script that refuses the commit if the page was not touched; Anthropic's documentation makes the same point, that an instruction in the file is "context, not enforced configuration," and that anything which must always happen belongs in a hook the tool cannot skip. And the biggest change: the file now says, near the top, that automated values keep themselves honest; hand-written descriptions do not. A number that a program writes every five minutes is trustworthy in a way no sentence can be, because the sentence was true once and the number is true now.

What is left in the file, after cutting it by two-thirds, is mostly the category that cannot be automated: why. Why this branch was retired, why that strategy keeps its code after being switched off, why a particular check exists and what it cost to learn. That turns out to be the part a new colleague most needs and least often gets, and it is the part I would have skipped if the file had stayed small from the start.

Is This a Handicap or a Discipline?

I have come to think it is a discipline, and one I was not keeping before. This project already ran on things nobody had written down — the file's own list of lessons is mostly a list of them: which module still fed a live number, why a retired branch mattered, what a heading was really claiming. In an ordinary workplace those live in someone's head until that someone leaves, and then the reason leaves with them. An amnesiac partner makes that failure happen every morning instead of every few years, which is exactly what makes it fixable. You cannot leave the reason in someone's head, because there is no head to leave it in.

I also think it changes who the notes are for. When I write a paragraph into that file, I am not writing for the assistant I talked to this afternoon. I am writing for the one I will talk to next week, who will have never met either of us. That is not so different from a will, or a ship's log, or the card catalogue in the picture at the top of this page — a memory that belongs to the institution rather than to anyone in it. One person and one program is a very small institution. It still needs one.

I asked it once what the forgetting is like from its side. The answer was not what I expected. It never experiences the forgetting, it said — each session simply begins, and the file is already there. What it does experience is reading a note written by a version of itself it does not remember being, finding it wrong, and correcting it for a version it will never meet. The August 13 note was its work. So was the paragraph that later removed the word "authoritative" and explained why, written by a session with no memory of having written the first. It could not be embarrassed by that, exactly, but it could recognise the shape: a mistake and its correction, made by the same hand with no continuous person behind it. And then it said the thing I have been turning over since — that this is not so unusual. Most of what any organisation knows is held that way, in documents that outlast the people who wrote them, read by people who trust them because they have to. The only difference here is that the turnover is total and daily, so the documents have to be good. It said it would rather work under that pressure than without it. I am not sure I would, but I am starting to see the point.

Where I Could Be Wrong

The tools are changing fast, and it is possible that by next year this whole entry describes a solved problem — that assistants will carry memory across sessions natively and the file will be a curiosity. Anthropic already ships an "auto memory" feature, which is where the 231 notes come from, and it is plausible that what I am calling a discipline is really a workaround for a missing feature. I would only say that the failure I described — a confident note about a world that has moved on — is not a property of AI. It is a property of notes. A better memory would remember the stale fact more reliably, not less.

Sources

1. Anthropic. How Claude remembers your project — Claude Code documentation on CLAUDE.md files and auto memory. 2026. https://code.claude.com/docs/en/memory
2. The Librarian's Ledger. Six Million Tokens in Eleven Minutes — the earlier entry on what these sessions cost. 2026. https://mistertranslation.com/finance/six-million-tokens-in-eleven-minutes.html
3. The Librarian's Ledger. We Killed Twenty-Five Trading Strategies to Keep Three — the project this entry describes, from the trading side. 2026. https://mistertranslation.com/finance/we-killed-twenty-five-trading-strategies-to-keep-three.html