What should I tell ChatGPT to remember about me?

Durable things you would otherwise retype: how you want to be answered, what you are working on, decisions you have already made, constraints that do not move, and what you have already tried and rejected. If knowing it would change the answer, it is worth saving.

What should I not save to AI memory?

Secrets and credentials, full chat transcripts, moods and one-off frustrations, and anything that changes week to week. Also leave out anything you would not want a different AI to bring up later, because memory does not know which conversation is which.

Why does my AI remember something that is no longer true?

Because most memory is append-only. You saved a fact, the fact changed, and nothing removed the old version. When something changes, say what changed rather than only stating the new fact, and delete the version that is out of date.

Is there such a thing as too much AI memory?

Yes. Recall gets noisier as the pile grows, and stale entries start contradicting fresh ones. Twenty good facts beat two hundred vague ones. In Tabula you can read, edit, and delete everything that is saved, which is what makes pruning possible.

What to Put in Your AI's Memory, and What to Leave Out

Nikhil BhimaFounderJuly 23, 20263 min read

A neat stack of blank cream cards in sunlight on linen, with many more cards scattered loose behind them, in warm film tones.

Everyone tells you to turn AI memory on. Nobody tells you what to put in it.

So most people's memory ends up as a junk drawer. Half a preference, a project that finished in March, one bad mood from a Tuesday.

More saved is not better. Twenty good facts beat two hundred vague ones, because the noise is what makes recall unreliable.

Here is how to decide.

Three questions

Before you save anything, ask:

Would you type it again? If you would re-explain it in a new chat, it belongs in memory. If you would not, it does not.

Does it change the answer? Good memory is load-bearing. If knowing the fact makes the output different, save it. If it is just trivia about you, it is noise.

Is it a fact or a mood? "I prefer short answers" is a fact. "This project is driving me mad" is a mood. Moods age badly, and they colour answers for months.

What earns its place

How you want to be answered. Tone, length, format, the words you never want used.
What you are working on. The project, where it stands, who it is for.
Decisions already made. So you stop relitigating them in every new chat.
Constraints that do not move. Dietary, accessibility, tools you cannot use, budget, timezone.
What you already tried and rejected.

That last one is the most underrated. Without it, your AI keeps proposing the thing you ruled out in March, and you keep explaining why not.

What to leave out

Anything secret. Keys, passwords, card numbers. Memory is readable by every AI you connect it to.
Transcripts. Save the conclusion, not the conversation.
Anything that changes weekly. It will be wrong before it is useful.
Anything you would not want a different AI to bring up. Memory does not know which chat is which.

The part nobody warns you about

Memory goes stale, and stale memory is worse than none.

Two things go wrong. Facts expire: you save "we are pre-launch", you launch, and six months later your AI is still writing pre-launch copy. And facts collide: "I am vegetarian" and "I do not eat mushrooms" look like the same fact to a naive system, so one quietly overwrites the other.

Three habits fix most of it.

Say what changed, not just what is true now. "We launched in June, so we are no longer pre-launch" is a better save than "we launched."

Be specific. "Prefers short answers" is weak. "Prefers answers under 150 words, no bullet lists" survives contact with reality.

Read your memory occasionally. If you cannot see what is saved, you cannot fix it. Whatever you use, insist on being able to read, edit, and delete it. That is why Tabula shows you everything it holds.

The short version

Save what you would otherwise have to repeat. Leave out what you would not want repeated.

And keep it somewhere you can actually look at.

Questions

What should I tell ChatGPT to remember about me?: Durable things you would otherwise retype: how you want to be answered, what you are working on, decisions you have already made, constraints that do not move, and what you have already tried and rejected. If knowing it would change the answer, it is worth saving.
What should I not save to AI memory?: Secrets and credentials, full chat transcripts, moods and one-off frustrations, and anything that changes week to week. Also leave out anything you would not want a different AI to bring up later, because memory does not know which conversation is which.
Why does my AI remember something that is no longer true?: Because most memory is append-only. You saved a fact, the fact changed, and nothing removed the old version. When something changes, say what changed rather than only stating the new fact, and delete the version that is out of date.
Is there such a thing as too much AI memory?: Yes. Recall gets noisier as the pile grows, and stale entries start contradicting fresh ones. Twenty good facts beat two hundred vague ones. In Tabula you can read, edit, and delete everything that is saved, which is what makes pruning possible.