A language model has no memory. What it "knows" during a conversation lives in its context window — and that starts out empty in every new session. For casual chatting, that's fine. But if you work with an AI on the same projects every day, you end up re-explaining every morning how things are done here, what went wrong last week, and why the database is named the way it is.
The remedy is unspectacular — and it ships with the tool: a folder of text files.
How it works
The AI that works on my websites is allowed to write into a notes directory. One file per insight, plus an index file with one line per note — which it reads at the start of every session, the way you'd glance at your own notebook in the morning. The notes come in kinds: who am I and how do I like to work? What instructions and corrections have I given — and why? What are we currently building? And reference knowledge: addresses, relationships, known traps.
Two properties make the system trustworthy. First: these are readable plain-text files on my machine. I can open, edit or delete any note — there is no hidden profile the AI compiles about me, just a notebook lying open on the table. Second: the folder is version-controlled. Every change is committed automatically at the end of a session — not because the AI "remembers to", but because the surrounding system enforces it. Reliability is better built into the environment than into good intentions.
What it buys: the end of repeat mistakes
The value shows most clearly with corrections. One example followed me for a long time: on my websites, changing pages moves keyboard focus to the main heading — an aid for people using screen readers. If you forget one particular style rule, some browsers draw a visible box around that heading. A tiny detail. I flagged it on every new website — including this one.
The last time, it became a note: what the mistake is, why it happens, how to avoid it from the start. It hasn't appeared since. That is exactly the difference between a correction and a stored correction: one lasts until the end of the conversation, the other beyond it.
By now the folder also holds less flattering entries — such as the warning about a misdiagnosis the AI once talked itself into. It reads its own accident reports before going back to work.
Where it goes stale
Notes carry an expiry date that isn't printed on them. "The configuration file lives in folder X" is true until someone tidies up. "Feature Y isn't finished" is true until it is. A stale note is more treacherous than no note, because it speaks with the authority of the written word — the same problem every human wiki has.
Three habits help: date the notes. Before acting on one, check that the world still looks as described — especially for file paths and system details. And delete what turned out to be wrong instead of only piling up new entries; a memory is only as good as its upkeep.
What to take away
The big chat providers are building memory features into their products too — there, the notebook lives with the provider. That can be convenient; you should simply know it exists and occasionally look at what's in it. The principle is the same as with my folder of text files: an AI with memory is more useful, an AI with inspectable memory is more trustworthy. Given the choice, pick the variant where you can read what has been written down about you.
/compact — the essentials, if context is running low:
A language model has no memory — the remedy is a folder of readable, version-controlled plain-text notes whose index the AI reads at the start of every session. A stored correction thus lasts beyond the single conversation and puts an end to repeat mistakes. Because a stale note is more treacherous than none, it helps to date notes, check them before acting on them and delete what turned out to be wrong — and an AI with inspectable memory is the more trustworthy choice.