What AI remembers about you

Sooner or later it happens to everyone. You open a fresh chat, ask a simple question, and the answer mentions your business by name, or your kids, or the price you quoted a customer three months ago. It feels like the AI has been getting to know you. It has not, at least not the way a person does. Something much simpler and much more controllable is going on, and once you see it you can decide exactly how much you want it to know.

The model does not learn from your chats

The model, the giant trained file that actually writes the answers, is frozen once training ends. How AI is trained covers that process; the short version is that chatting with it does not change it. Your conversation tomorrow runs on the exact same model as your conversation today.

What changes is the app, the program wrapped around the model (ChatGPT, Claude, or Gemini on your phone or in your browser). The app can keep notes about you, search your old conversations, and hand the useful bits to the model each time you send a message. That is what "memory" means in these tools: a filing system the app runs on your behalf, not a brain that grew.

Using your chats to train a future model is a different thing from memory, with a different switch. It gets its own section below.

Five layers of "remembering"

When people say "the AI remembers," they usually mean one of five separate features. They look alike from the outside because they all end the same way: more text quietly added to what the model reads. The difference is who wrote that text and how long it sticks around.

• This chat. Everything said in the current conversation. This is the context window, the model's short-term working space, and it empties when you start a new chat.
• Saved memory. Short notes the app writes about you, either because you said "remember this" or because it decided a detail was worth keeping. You can read them as a list.
• Past chats. The app searches your old conversations for something relevant and pulls in a few pieces. Nothing is saved as a separate note; it goes back to the originals.
• Projects. A folder with instructions and files you set up on purpose for one kind of work. Projects and knowledge covers them in full.
• Custom instructions. Standing rules you type once in settings that ride along in every chat. Setting it up in your tool walks through writing good ones.

The last two you wrote yourself, on purpose. The middle two the app wrote for you, and that is where the surprises come from. Tap through the layers to see what each one actually hands the model.

Where to see it, edit it, and delete it, as of October 2026

Menu names move around, so treat these as directions, not a map. This is what each company's own help pages said on October 4, 2026.

• ChatGPT. Go to Settings, then Personalization, then Memory. Depending on your account you will see switches named "Reference saved memories" and "Reference chat history," plus a way to manage the saved list, where you can delete one memory or clear them all. Turning off saved memories also turns off chat history reference. OpenAI says that when you turn chat history reference off, what it remembered from past chats is scheduled for deletion within 30 days, but the chats themselves stay unless you delete them. The reverse matters too: deleting a chat does not necessarily delete a saved memory made from it. To fully remove something, delete both.
• Claude. Go to Settings, then Memory. The switch "Generate memory from chats" controls saved memory, and "Search and reference chats" controls whether Claude can look through past conversations. What Claude remembers is listed under Topics; open one to read it, edit it, or delete it. Anthropic says memory is on by default on Free, Pro, and Max, while searching past chats is a paid plan feature. Each project keeps its own separate memory, so a client project does not leak into your personal chats.
• Gemini. On the web, open Settings & help, then Personal Intelligence. Memory is a switch there, and "Instructions for Gemini" (called Saved info in some regions) lives in the same menu. Google says Memory needs you to be 18 or older, on a personal Google account, with Keep Activity turned on, and that it does not work inside Gems or Live chats. To remove what it learned, delete the chat in your Gemini activity at myactivity.google.com, or correct it directly in a chat. Google notes there can be a short delay before a deleted chat stops being used.

A practical habit: you can ask any of the three, in plain words, "Did you use anything from past chats or memory for that answer?" It is not a perfect audit, but it often names the note it used, which tells you exactly what to go delete.

Temporary and incognito chats

As of October 2026, every major app has a way to talk without leaving a trail in your memory. It is the right tool for a one-off question you do not want following you around: a medical worry, a gift for your spouse, a rough draft of a hard email to an employee.

• ChatGPT Temporary Chat does not create or update memories. If you choose to save a temporary chat, it becomes a normal chat and follows your usual settings again.
• Claude incognito chats are not saved to memory or to your chat history, and Claude will not reference them later. Claude also lets you switch memory off for a single chat before you send the first message.
• Gemini Temporary Chat is Google's own suggestion for anything you do not want Gemini to remember or use to personalize your experience.

Temporary does not mean invisible. The company still keeps a copy for a while for safety reasons; the privacy lesson has those retention numbers. What it means is that this conversation will not come back to shape future answers.

Training is a separate switch

Memory decides what the app hands the model about you. Training decides whether your chats become practice examples for the next model the company builds. They are controlled in different places, and turning one off does not turn off the other. The full list of training switches, and what they do and do not undo, is in Privacy for your files; it is worth five minutes if you have not done it.

One wrinkle as of October 2026: in Gemini the two are linked. Memory only works with Keep Activity on, and Keep Activity is also the setting under which Google can use a sample of your uploads, with human reviewers, to improve its services. In ChatGPT and Claude, memory and training sit in different settings menus, each with its own switch.

So it is two questions, not one: should this app remember me, and should my chats train the next model? Yes and no is a common, sensible pair of answers.

Try this yourself

Open the chat app you use most and find its memory page using the paths above. Read every entry. You will probably find at least one thing that is out of date, and possibly one you did not expect it to keep. Delete or fix anything wrong. Then open a normal new chat and paste this:

What do you know about me from saved memory or past chats? List each item and tell me where it came from.

Now open a temporary or incognito chat in the same app and paste the exact same question. Compare the two answers side by side. The gap between them is your memory layer, made visible. If the first answer surprised you, you now know where the switch is.