Part of the answer is a list you can go and read in about a minute. The rest of it is a claim worth taking apart, because a mirror and a witness are not the same thing.
Two different questions are hiding inside this one, and they come apart fast.
The first is literal. What has ChatGPT actually written down about me, where is it kept, and can I read it? That has a real answer, it takes about a minute, and most of what ranks for this search skips straight past it in favor of clever prompts. The second question is the one people are really asking: does it know me? That one deserves a slower answer, and the honest version of it is not flattering to the tool.
Do them in order. The list first, then the prompts, then the ceiling.
What it actually holds
Four separate piles, and it is worth knowing which is which, because they behave differently and you delete them differently.
Saved memories. Short written notes it kept from things you said. This is the only pile that exists as a browsable list, and it is the one people mean when they say ChatGPT remembers them.
Reference to past chats. A setting, not a list. It lets ChatGPT draw on conversations you have already had. You cannot open it and read it. You only notice it when it brings something up.
What you told it on purpose. Your custom instructions: your name, what you do, what you want it to know about you, the tone you asked it to take. Easy to set once and forget, and it colors every answer you get afterward.
Your account and chat history. The raw transcripts and the account data. Much larger than the memory list, and it does not live in the memory list.
One thing is true of all four: none of it exists unless you put it there. Hold on to that, because it is the whole argument later on.
How to go and read it
Settings, then Personalization, then Memory, then Manage. That screen is the list. Every entry is a note it wrote down about you. You can delete any single one, and there is a Clear all if you want the lot gone. Reading it top to bottom is the single most useful thing on this page, and almost nobody does it.
Same Memory panel, two toggles. Reference saved memories and Reference chat history. They work independently, so you can leave one on and kill the other. Turn both off and it starts every conversation with no continuity at all.
Settings, then Personalization, then Custom instructions. This is what you handed it deliberately. Worth rereading even if you never touch it again, because a description you wrote of yourself eighteen months ago is quietly shaping today's answers.
Settings, then Data controls, then Export data. Confirm, and it emails you a link when the file is ready. Expect to wait, and download it quickly, because the link expires about a day after it lands. This is the big one: your actual chat history, not a summary of it.
Temporary Chat, the control near the model picker, is how you ask something without it reading from or writing to any of the above.
Two honest caveats. Menu labels move around between updates and between the app and the web, so if a name here does not match your screen, open Settings and look under Personalization first and Data controls second. And the memory list is not the whole picture, because the past-chats setting has no list view. What you can audit and what it can use are different sizes.
The prompts that are actually worth running
Every article on this query hands you five prompts. Most of them return a warm paragraph. These five are ordered by how much they can tell you that you did not already hand over.
The audit. "List back everything you have saved about me, exactly as it is stored, and add nothing." Boring and the most useful. It is the closest thing to reading the file out loud.
The inference. "What do you believe about me that I never told you directly, and what did you infer it from?" Forcing it to cite the source is what stops it inventing.
The contradiction. "Where have I contradicted myself with you? Quote both sides." This one is the best on the list, and I will say why in a second.
The absence. "What do I bring up most often, and what have I never once mentioned?" The second half is the part that stings.
The outside view. "Describe me to someone deciding whether to work with me. No compliments." Removing the compliments is doing most of the work.
The contradiction and the absence prompts beat the other three for a structural reason. A summary of what you said is still what you said, rearranged. But a contradiction is a relationship between two things you said, and an absence is a shape in the gaps. Neither is something you can hand over on purpose, which makes them the only two answers that are not just you, tidied.
The "chatgpt tell me about myself" prompt
This is the phrasing people search for by name, and asked flat it is the weakest prompt of the set. Two forces work against you. The model leans agreeable, because being pleasant to use is part of the job. And the raw material is your own account of yourself, so the flattery is being applied to a version of you that was already edited before it arrived.
You get a much better answer by taking both away. Paste this into a fresh chat:
Using only my saved memories, my custom instructions, and our past conversations, tell me about myself. Rules: no compliments, no encouragement, no summary of my strengths. Quote or paraphrase the specific things I said that you are drawing on. Keep what I told you directly separate from what you inferred. Finish with the subject I circle and never land on.
Then read only the inferred half. The direct half is you handing yourself back to yourself. The inferred half is the only part doing work, and even there, hold it loosely: a description can feel devastatingly accurate and still fit almost anyone. That effect is strong enough that we built a small test that lets you catch yourself falling for it, and it is worth two minutes before you trust any profile of yourself, including the one you just generated.
The ceiling nobody mentions
Everything in that list is there because you typed it. That makes ChatGPT an unusually good mirror. It also means it inherits every limit a mirror has.
It only saw the days you opened it. And you do not open it at random. You go to it in particular moods, for particular tasks, in a particular register. Whole seasons of your life are simply not in the file, and the missing part is not a random sample.
It has your account of events, not the events. By the time a thing reaches the chat box it has already been through the part of you that decides how to tell it. The narrator is the subject.
It has no idea how you land on anyone else. It has never watched you in a room. The outside data problem is the oldest one in self-knowledge, and a chat log does not touch it.
It agrees with you. An instrument built to be pleasant is a poor witness. If it has never told you something you did not want to hear, that is not evidence you are doing well.
Nothing in it is dated in a way you can use. A memory list is a pile, not a record. It can describe you. It cannot show you that you moved, which is the thing most people are actually asking about when they ask what it knows.
None of that makes it useless. It makes it one source with a known bias, which is honestly the most you can say about any self-report, including every "who am I" test ever built. The mistake is treating a very good mirror as though it were an outside read.
An outside read, in one paste.
Paste something you wrote somewhere ChatGPT never saw. An old journal entry, a long message, an email you sent. It tells you how you land on a reader and the habit you cannot hear in your own voice. Free, no signup.
0 characters
Nothing you paste is stored, and no account is created.
How you land
The tell
That came from one sample, and it is still one sample. A reading works from six questions and then keeps the answer where you can watch it move.
That is one of four free tools. No signup, no email, no card on any of them, and nothing you paste is written down. The other three name one word for you out of a paragraph, compare something old you wrote against something recent, and let you catch yourself believing a profile that was written for nobody in particular.
The thing a chat log structurally cannot do
Ask ChatGPT whether you have changed and it will give you a confident answer built out of whatever you happened to mention. It has no fixed point to measure against. It cannot tell the difference between you changing and you being in a different mood the week you typed.
That is the same failure a personality test has, arriving from the opposite direction. A test freezes you on the day you took it. A chat log is a heap of days with no spine running through them. Neither one can show you movement, and movement is what you were asking about.
What we built instead is a record. Six questions, about eight minutes, and it returns one word and a card dated the day the record opened. The back of the card carries the evidence: beats on record, weeks active, and a "verified continuous since" date that can only ever get older. Then the record grows as you add to it, compresses into something shorter and truer when enough has piled up, and grows again from there. It never resets you to a type, because there is no type to reset to.
The difference from a memory list is time. Nobody can fake time, and nobody can catch up to a record that started before theirs. The full walkthrough shows what the record holds and how it moves. The word and the first reading are free, and what the ongoing room costs is on one page with no asterisks.
Common questions
Can I see everything ChatGPT knows about me?
You can see the saved list. Open Settings, then Personalization, then Memory, then Manage, and every stored entry is there to read or delete. You cannot browse the other half: the setting that lets ChatGPT draw on past conversations has no list view, so it surfaces only when it mentions something. For the full pile of raw chats and account data, use Settings, then Data controls, then Export data.
How do I delete what ChatGPT knows about me?
In Settings, Personalization, Memory, Manage, each saved entry can be deleted on its own, and Clear all wipes every entry at once. In the same Memory panel two independent toggles, Reference saved memories and Reference chat history, stop it drawing on any of it. Turning both off leaves no continuity between chats. Temporary Chat, near the model picker, neither reads from nor writes to memory.
What is the best prompt to ask ChatGPT what it knows about me?
Two beat the rest. Ask where you have contradicted yourself and make it quote both sides, and ask what you have never once mentioned. Those are the only prompts whose answers are not just a tidier version of what you already claimed, because a contradiction and an absence are structural. A flat request to describe you returns a summary of your own account of yourself.
Can ChatGPT tell me something about myself I do not already know?
It can spot patterns across more of your own words than you can hold in your head, which is genuinely useful. It cannot get outside your account of yourself. Everything it holds arrived because you typed it, on days you chose to open it, in the version you were willing to tell. It also has no dates you can measure against, so it can show you a picture but never movement.
Read the memory list first. Then, if you want the half a chat log cannot reach, six questions and about eight minutes, no card.