The analogy

Think of the cafe down the street where the barista, after you've been going for a while, makes your coffee the way you like it without asking. It's not magic or affection: somewhere, in their head or on a notepad, they have a line, "the one at the back table, cold macchiato, little sugar." When you walk in, they reread that line and act accordingly.

The AI's long-term memory is that notepad. It's not that its brain has changed for you: there are lines written about you, on a sheet tied to your account, that get reread every time you open a conversation. That's all. And like the barista's notepad, you can flip through it, correct a wrong line or tear it out.

How it really works

On ChatGPT, memory works in two ways. There are the explicit notes, the ones you ask for ("remember I'm celiac"), and there's the reference to past conversations, that is, cues the system draws on its own from previous chats to make the next ones more relevant. At the start of a new chat this material is placed in front of the model again, which starts off already "informed" about you. Over time the notes update: a "I'm leaving for Lisbon in July" becomes "was in Lisbon in July" once the date passes.

Gemini and Claude offer similar personalization mechanisms tied to history or projects. The principle doesn't change: it's text about you, kept separately and reread, not a model that has learned your person.

What you can do in practice

  • Tell it explicitly what you want it to keep: "remember I write for an audience of beginners." Direct notes are the most reliable.
  • Browse and manage the memory from the settings: there you see what it noted, correct the wrong notes and delete the ones you don't want.
  • For a request that must leave no trace, use the temporary chat: it doesn't read the memory and doesn't create any.
  • If you want to start clean, or for privacy, you can switch the memory off entirely.
  • Avoid letting it memorize sensitive data (detailed health, passwords, others' data): they stay written until you delete them.

A common misconception

People think that "having memory" means the model has learned you, that it has become a bit yours. It's not so: the model is identical for all users in the world. What changes is only the note sheet about you, reread at the start of your chats. And it doesn't remember everything you've said: it remembers what was noted. Confusing the two leads you to expect it to "already know" things it never put in the notepad.

Frequently asked questions

Where do I see what the AI knows about me?

In the account settings, in the section dedicated to memory or personalization. There you find the list of notes, you can edit them one by one or clear them all at once. If the list surprises you, it's time to clean it up.

Is it private and secure?

The notes are tied to your account and you control them, but they remain data kept on a server: treat them as such. Don't let it memorize what you wouldn't write on a note that might not stay only yours, and do a cleanup every so often.

Can I switch it off?

Yes. You can turn off the explicit notes, the reference to past conversations, or both. With memory off, every chat starts from scratch, like in the old days: less convenience, more privacy. The choice is yours.

If I delete a chat, do I also delete what the AI has learned about me?

No, and that's the misunderstanding that leaves more data lying around than expected. Long-term memory is separate from the conversations: a note like "lives in Bologna" survives even if you delete the chat where you said it. To really remove it you have to go into the memory management and delete it there. Clearing the chat history and clearing the memory are two different operations.