AI memory and chat history: what the assistant keeps about you

An assistant that seems to remember you is really running three separate systems, and mixing them up is why people cannot work out what it knows. The context window is the current conversation, held only for as long as that conversation runs. Chat history is the stored list of past conversations, which sits on a server and is normally not read by the model at all. Memory is a short set of facts the assistant has saved about you and quietly adds to every new chat. Deleting one of these does not touch the others, and each has its own setting in a different part of the interface.

The three layers side by side

The model itself remembers nothing between requests. It is a function: text in, text out. Everything that looks like continuity is the application putting words back in front of it.

The context window is not memory

Within one conversation the assistant appears to recall what you said ten messages ago because the entire transcript is resent with every turn. Nothing is stored in the model. When the conversation grows past the window, the oldest material is dropped or compressed into a summary, which is why a long chat starts forgetting its own beginning and why details you gave early can silently vanish. The mechanics and the practical limits are set out in why an AI forgets the start of a long chat.

This has one useful consequence. If an assistant is stuck on a wrong assumption, a new conversation genuinely clears it, as long as no saved memory is carrying the assumption forward.

Chat history is a list, not a memory

The sidebar of past conversations is storage for your benefit. By default the model does not consult it: opening a new chat gives you a blank slate even though last week's conversation is sitting three inches to the left. History exists so you can find things again, and so the provider can meet its own retention and safety obligations.

Some assistants now add a retrieval step that searches past chats and pastes relevant fragments into the new conversation. That is worth understanding as a search, not as recall. It can drag in an old conversation that is no longer accurate, and it means the boundary between chats is thinner than it looks. If the feature exists, it has its own switch, usually beside the memory controls.

Temporary or incognito chat modes are the opposite: the conversation is not written to history and does not update memory. It still travels to the provider and is still retained for a period for abuse monitoring, since the journey described in where your prompts actually go is unchanged by a client side setting.

Memory is a short file of facts about you

Memory is smaller and simpler than most people imagine. It is a list of plain sentences: you work in logistics, you have two children, you prefer concise answers, you are learning Portuguese. Entries arrive two ways. You can add one deliberately by saying remember this, or the assistant can decide on its own that something is worth keeping. Those saved lines are then inserted into the hidden instructions at the start of every new conversation, which is why they cost a little of your context on every chat, as explained in what tokens are.

Because it is a list of sentences, you can usually open it, read every entry and delete individual ones. Do that at least once. The things that go wrong are predictable.

• Wrong inferences. You research a medical condition for a relative and the assistant records it as yours. You ask about divorce law for a friend and the entry stays for a year.
• Stale facts. Old job, old city, a project that ended, a preference you have changed. The assistant has no way to know the world moved on.
• Context leakage. Memory does not distinguish work from personal. A saved detail about your health can surface in a conversation you are screen sharing in a meeting.
• Poisoning. If the assistant browses the web or reads documents for you, text inside a page can try to instruct it, and a saved false memory persists long after the page is closed. This is rare, but it is the reason to read the list rather than trust it.

Shared links, files and connected apps

Three side channels hold conversation content and none of them live in the history list.

Shared links. Sharing a chat creates a snapshot at a public web address. Anyone with the link can read it, links have been picked up by search engines in the past, and deleting the original conversation does not always revoke the link. Shared links have their own management screen, and it is worth emptying it occasionally.

Uploaded files. Documents, spreadsheets and images are usually stored separately from the text of the chat, sometimes in a workspace that outlives the conversation that used them. Clearing your history may leave the files behind.

Connected apps. Granting an assistant access to your mail, calendar or cloud drive means it can pull real content into a conversation on request, and that retrieved content then sits in the conversation log like anything else you typed. Review these grants the way you would review any account permission, using the same reasoning as closing an account properly: access you no longer use is access you should revoke.

What deleting each layer actually does

Deleting a conversation removes it from your history and starts a purge clock on the provider's systems, commonly around thirty days, after which it should be gone from live storage. Backups and anything retained for a legal or safety reason can outlast that. Deleting a memory entry stops it being injected into future chats. Neither action reaches a model that was already trained on the text, because training is not reversible and there is no thread connecting a finished model to your sentence.

That ordering matters. Turn training off before you paste something sensitive rather than after, which is a forward looking decision covered in how to stop your chats being used for training. The full mechanics of what deletion does and does not reach are in how to delete your AI chat data, and the wider picture of what a provider keeps is in what a chatbot does with the personal data you type.

A ten minute review worth repeating

Put this in the calendar quarterly. It is faster than it looks and it usually turns up something you would rather not have stored.

1. Open the memory or personalization list and read every entry. Delete anything stale, anything about another person, and anything you would not want on a screen behind you.
2. Check your custom instructions. People write them once and forget the job title in there is two roles out of date.
3. Open the shared links screen and delete links you no longer need public.
4. Review connected apps and revoke anything you have not used recently.
5. Confirm the training setting is where you want it, and note that it applies going forward only.
6. Check which account you are signed in with for work, since a personal login carries none of your employer's contractual protections.

Two habits do more than any setting. Use a separate account, or at least temporary chats, for anything you do not want shaping future answers. And when an assistant says something oddly personal or oddly wrong, ask it what it remembers about you, then go and look at the list yourself.