Can I just ask it to remember?

No, not by asking inside the conversation. It will politely agree, then forget the moment you close. The only reliable memory is a file on your disk that it re-reads at startup.

Do I have to write everything into that file?

No, and doing so backfires. The official documentation recommends staying under 200 lines. An overlong file is followed less closely than a short, precise one.

Does this work with every AI?

The principle is the same everywhere. An instruction file is read at startup, and only its name changes from one tool to the next.

Why your AI forgets your instructions every single time

You explain your project. It understands, it works well. You close the window. The next day it knows nothing: not what your site does, not what you forbade the day before. You start the same explanations over every morning, and you are finding it exhausting.

Your AI remembers nothing between sessions: every conversation starts with a blank memory. It only keeps what is written in an instruction file that it re-reads each time it starts. Without that file, you explain your project again every single morning, and it still drifts.

It isn’t a bug, it’s how the thing is built

You explained your project on Monday. On Tuesday it starts from nothing. You explain again. Wednesday, same story.

This is not stubbornness and nothing is broken. The official documentation says it plainly: every session starts with a blank memory. What you typed yesterday no longer exists. What you type today will be gone tonight.

A coding AI works with whatever sits in front of it right now. That workspace is large, but it empties completely when you close the window.

The only thing that survives closing

There is a mechanism built for exactly this: an instruction file placed inside your project, which the AI reads automatically when it starts.

This is not a workaround. It is the official way. The file holds whatever you keep repeating: what the project does, how you want to be spoken to, what is off limits, what must never be touched.

Written once, read every time.

Length matters more than you would think

The natural instinct, once you discover this file, is to pour everything you know into it. That backfires.

The official documentation recommends staying under 200 lines, with a hard ceiling of 40,000 characters. The reason is simple: the longer the file, the more room it eats in the AI’s workspace, and the less closely it is followed.

A short, precise file is respected more than an exhaustive wall of text. It is counter-intuitive, and it is measurable.

What belongs in it

Three things, in this order.

What the project does, in three lines. Not a sales pitch, just enough for it to know where it has landed.

How you want to work. Does it draft a plan before changing anything? Does it ask before deleting? Does it explain, or does it just execute?

What is forbidden. This is the part everyone skips and the part that pays off most. One written prohibition saves hours of repair work.

Why “I already told you” does not work

It is the sentence everyone ends up typing, often in capitals. It never has the effect you expect, and the reason has nothing to do with goodwill.

Within the same conversation, the tool does have what you wrote earlier in front of it. Repeating it teaches nothing new — at best it brings the instruction back into focus for the next reply. And when you close, that reminder disappears with everything else.

In a new conversation it is starker still: the sentence points at something that does not exist. There is no “already told”. There is only what is written in a file, and what you just typed.

Hence the reflex that changes everything: the second time you type the same explanation, it is not a repetition, it is a signal. That sentence belongs in the file, not in the conversation.

What happens when the conversation fills up

There is a second kind of forgetting, less well known, that strikes inside a single conversation.

The AI’s workspace — its context window — is finite. Everything you exchange piles into it: your messages, its replies, the files it opened, the dead ends. While there is room, all of it is re-read on every reply. When it overflows, the oldest goes first.

What that looks like in practice: an instruction given at the start of a session can quietly stop being followed two hours later, with no error, no warning. It feels like it has “stopped listening”. In reality your instruction has dropped out of its field of view.

An instruction file does not suffer from this, precisely because it is re-read at startup rather than carried by the conversation.

How to tell whether it worked

Close everything. Reopen. Ask it what your project does.

If it answers correctly without you re-explaining anything, the file is being read. If it asks again, the file is in the wrong place or it is not being loaded, and that is the thing to fix before anything else.

Run this test again after every change of folder or machine. It takes ten seconds, and it is the only proof that counts: until you have done it, you are assuming.