How to do it
The AI "knows" about you on two levels: the facts it has saved on purpose (the memory) and what it can infer from the open conversations. The first you control and delete; the second disappears when you close or delete the chats.
Open the memory. It's the archive of the facts the system notes about you from one chat to another. On ChatGPT it's in the data controls or in personalization; on Claude and Gemini in the settings dedicated to memory. There you find the explicit list: names, projects, preferences it has recorded.
Ask it out loud, in the box. Besides the list, have the AI itself tell you what it remembers. Sometimes it's clearer than the technical list, and it surfaces inferences you didn't expect.
Check the custom instructions. They're the notes you've given ("answer me in English", "I'm a lawyer"): they count as things the AI knows about you and enter every answer. Review them.
Scroll through the history. The open conversations are the context the AI can draw on. What's no longer in the history is no longer within reach of a new chat.
Delete what you don't want. From the memory list, remove individual entries or wipe everything; correct the custom instructions; delete the chats you'd rather not leave.
If you can't find the memory section because the interface has changed, ask the AI: "Where do I see and manage what you've memorized about me?".
The working syntax for the direct check:
List everything you've memorized about me persistently,
item by item, as you would see it in the memory settings.
Then tell me what you're inferring about me just from this conversation.
Clearly distinguish the two things and don't add made-up assumptions.
A concrete example
Marco suspects ChatGPT "knows" him too well. He opens the memory and finds three entries: his job, his city, the name of a work project. Then he asks in the box "What have you memorized about me?" and the AI confirms them, adding that from the current conversation it's inferring he's house-hunting. Marco deletes the work project entry (confidential), keeps the job and city (they're handy), and for the house search he opens a temporary chat so nothing is left.
When it does NOT work (and how to fix it)
If the AI says it doesn't remember anything but then seems to remember you
It can draw on the history of the open chats even if the explicit memory is empty. Fix: to start truly clean, use a temporary chat or delete the old conversations; that way there's no more context to fish from.
If the memory list and the AI's answer don't match
The list in the settings is the reliable source; the answer in the box is a reconstruction and can be wrong or made up. Fix: trust the settings list to decide what to delete, use the spoken answer only as a cue to check more carefully.
If you've deleted the memory but fear something remains somewhere
Deleting the memory empties the visible archive, but data already used in a past training round can't be recovered or inspected. Fix: for the future turn off training; for the past there's no possible check, so prevention is worth more than checking.
A tip from someone who actually uses it
Do this check every couple of months, the way you empty a mailbox. The memory fills up on its own, silently, with every conversation: a name here, a detail there. A periodic look at the list spares you from discovering one day that the system has noted things you assumed were forgotten.
Frequently asked questions
Does deleting a chat also delete what the AI learned from that chat?
Not necessarily. If from that conversation the system saved a fact into the memory, that fact remains even after you delete the chat. You have to empty the memory separately: deleting the conversation removes the text, not the extract that was saved from it.
Does the AI profile me the way social media do?
It's the belief to clear up. An AI assistant with memory saves facts useful to continue the work, it doesn't build an advertising profile to resell the way social platforms do. But the principle of caution remains: everything it saves is material about you, so checking and pruning it is a good habit, exactly like reviewing app permissions.