ChatGPT Memory: What It Remembers About You, How To Fix It, And How To Turn It Off

ChatGPT's memory used to be a list. You told it something, it wrote it down, and the list slowly filled up with stale facts like "you're going to Singapore in July" long after July was over. In June OpenAI rebuilt the whole thing. It now watches your chats, your files and your connected apps in the background, writes itself a running summary of who you are, and quietly updates that summary when something changes. It is genuinely better, and it is also the reason more of your life is sitting in there than you think.

Here's the full tour: where to read what it knows, how to correct it in ten seconds, how to teach it on purpose, and the switches that decide when it is allowed to remember you at all. Ten minutes, and you will have a ChatGPT that actually knows you, minus the parts you never meant to tell it.

Go to Settings, then Personalization, then Memory, and open the Memory summary with Manage. That is the page most people have never seen. It is a few paragraphs in plain English about your work, your hobbies, your preferences, your projects, with a "last updated" stamp at the top, and it is what ChatGPT reads before it answers you.

Then ask it directly, because the summary is a snapshot and the memory underneath is broader than what fits on the page.

Tell me everything you remember about me, organised by topic: who I am, my work, my family and the people I mention, my preferences, my current projects, and anything else you have picked up. Be specific and include the small things. Then flag anything that looks out of date, anything that was probably a one-off rather than a fact about me, and anything you are unsure about.

Read the answer slowly. Almost everyone finds at least one thing they never meant to keep, and one thing that is simply wrong.

Three ways, all on the summary page. Type a correction into the text box at the bottom and it updates. Highlight any line to correct that one thing. Or open the three-dot menu and choose Refresh to make it rewrite the summary from scratch. And in any normal chat you can just say it.

You have a memory that [THE WRONG OR OLD THING]. That is no longer true. The current version is [THE RIGHT THING]. Update your memory, and from now on do not bring up the old version.

For a proper clean-out, do it in one go instead of one fact at a time.

Go through everything you remember about me and give me a list of memories that are probably stale: things older than six months, one-off requests that got saved as if they were permanent, and anything that contradicts something newer. Number them. I will tell you which ones to forget and which to keep.

Deleting a chat does not delete a memory, and deleting a memory does not delete the chat. They are separate, which surprises people. And if you turn memory off completely (three-dot menu, Delete and turn off memory), your past chats stay exactly where they are.

Memory works best when you give it preferences rather than facts. A fact ("my sister is called Anna") is nice. A preference ("never use bullet points when you write an email for me") changes every answer from now on. Sit down once and tell it how you want things done.

Remember these as permanent preferences and apply them every time without me asking. How I write: [SHORT, WARM, NO EXCLAMATION MARKS, WHATEVER FITS YOU]. How I want answers: [THE FORMAT YOU LIKE, HOW LONG, HOW DIRECT]. My situation: [ONE LINE ON YOUR WORK], [ONE LINE ON YOUR LIFE]. Things to always ask before doing: [ANYTHING YOU WANT CHECKED FIRST]. Confirm what you saved.

Custom Instructions (in the same Personalization settings) still exist and still apply. The difference is that you write those by hand and they never change, while memory updates itself. I use Custom Instructions for the rules that must never drift and memory for everything else.

There are three switches, and knowing them is the difference between memory being useful and memory being a liability.

• Temporary chat. Open a new chat and tap the pill-shaped Temporary button in the top right. Since August you get a choice: personalized, which can read your memory but never writes to it, or non-personalized, which uses none of your memory or custom instructions at all. One of them is preselected, so check which before you type, because you cannot switch once the chat starts. Nothing from a temporary chat is saved to memory or history unless you save it.
• Project-only memory. Inside any project, open the three-dot menu, then Project settings, then Memory, and choose Project-only. Chats inside the project remember each other but never touch your main memory, and your main memory never leaks in. Perfect for a job hunt, a medical thing, or anything you want walled off. You can change it on an existing project now, and shared projects are always project-only.
• The training switch. Memory is one setting. Whether your chats and memories can be used to train OpenAI's models is a different one, under Settings, then Data controls, then "Improve the model for everyone." Turning that off does not touch your memory. Turning memory off does not touch that. If you want both, flip both.

Every plan has all three. Free gets a shorter memory than Plus, but the controls are identical.

The Honest Part

A memory that updates itself can also be wrong on its own, and a wrong memory quietly shapes every answer after it, which is why Step 1 is worth doing once a month rather than once ever. The summary page is a partial view, so if you correct something and it seems to come back, ask it directly and correct it in the chat too. And if two people share one login, the memory blends you into one person, so that is the account to keep memory off on.

Open the summary page now. Whatever is in there has been steering your answers for months. Ten minutes fixes it.