Skip to content
CSuite
Nov 19, 20269 min read

Does your AI remember you?

That “memory updated” notice means the app saved a note about you. You can read it, fix it and delete it.

On this page
One new chat · what the model is handed
You type one line. The app adds its notes about you before the model reads it.
What you typed

“Suggest a quick dinner for tonight.”

The reply suggests a meat-free dish you can cook after a late shift, and it feels like the assistant knows you.

What the model received
Notes saved from earlier chats
  • - Works night shifts as a nurse
  • - Vegetarian
  • - Prefers metric units
  • - Planning a trip to Lisbon in March
Suggest a quick dinner for tonight.
A drawn example with invented notes, not a real account. The mechanism follows the vendors’ own documentation, read on November 19, 2026.

You mention once that you are vegetarian. Three weeks later, in a new chat about something else, the assistant plans a dinner with no meat in it. A small grey line may have flashed past at the time: “Memory updated.” It is fair to wonder what was written down, where it lives, and whether the AI has learned you.

It has not learned you. The app saved a short note, and it shows that note to the model each time you start a chat. That is good news, because a note is something you can read, correct, move and delete. This post explains how the notes work in the main assistants and how to check yours.

Memory is a set of notes, and the model itself does not change

A language model is a very large file of numbers. Those numbers are set during training and stay fixed while you chat. The model that answers you today is the same file that answers everyone else, and nothing you type rewrites it. We covered that sealed box in why your AI doesn’t know what happened last week.

So memory has to live somewhere else. The app keeps a few lines of text about you in an ordinary database. When you send a message, it places those lines in front of your words, and the model reads the whole thing as one page. The model is like a new temp who is handed your file every morning: helpful because of the file, with no recollection of yesterday.

How a memory is made and used
1. Write
The app, or you

A short line of text is saved when you say “remember this”, or when the app decides a detail is worth keeping.

2. Keep
A database or a file

The line sits beside your account, outside the model. The model’s own numbers do not change.

3. Read back
The app

When you send a message, the saved lines are placed in front of it, and the model reads both together.

Nothing in these three steps retrains the model. Remove the saved line and the next chat starts without it.

The documentation says this plainly. The open-source app Open WebUI “adds the user’s stored memories to the model’s context so the model is aware of them”. Anthropic describes the memory tool it offers developers as a way to store and retrieve information “in a directory of memory files”. Microsoft says what Copilot infers from your chats “becomes memories that you can manage”. In every case the memory is text, saved outside the model.

An assistant’s memory is closer to a card file than to a brain. Cards go in, and the same cards are pulled out for the next conversation. Illustration generated with Nano Banana 2.1 via Runware.

There are two kinds: notes about you, and a search of old chats

The first kind is a list of facts. ChatGPT calls them saved memories. Claude stores memory “as a set of individual topics as you chat”. Gemini keeps a memory summary, and a separate box of standing instructions that Google’s privacy notice calls “Instructions for Gemini” (“Saved Info” in some places). These lists are short, and they come along to every chat.

The second kind is a search. Instead of carrying a list, the app looks through your old conversations and pulls out passages that seem related. ChatGPT has a setting for referencing chat history. Claude can be prompted to search your previous conversations, a feature for paid plans. Gemini offers personalization “using memory of your past chats”.

The difference matters when you tidy up. A list can be read from top to bottom, so you know what is in it. A search has no list. What it finds depends on what you ask, and the only way to remove something from it is to remove the conversation it came from.

Both kinds also compete for room. Everything handed to the model has to fit in its context window, the page it can read at once, and long pages are read less carefully, as we showed in your million-token context window is lying to you. That is why memory stays small. Open WebUI caps the memories it injects at 2,000 characters by default, which is roughly 300 words. A few hundred words is enough to hold your job, your diet and your current project, and not much more.

Every app lets you read the notes, in a different place

Because the notes are text, every major assistant now has a page where you can read them. The table below shows where to look, and the one detail in each app that surprises people.

Where five assistants keep their notes about you
From each maker’s help pages, read on November 19, 2026. Menu names change often, so look for the word “memory” in settings.
ChatGPT
The saved memories list in settings
Deleting a chat leaves the saved memory in place. Delete the memory as well.
Claude
Settings, Your Claude, then Manage next to Your memory. Entries are listed as Topics.
Deleting a chat does not remove entries made from it. Each project has its own memory.
Gemini
The memory summary in settings, or ask “What do you know about me?”
Works the other way round: you delete every chat that holds the detail.
Copilot
Chat settings, Personalization, then Manage saved memories
Turning memory off does not delete what is already saved.
Open WebUI
Settings, Personalization, then Memory
Runs on your own computer, and the notes sit in its local database.
Sources: OpenAI’s Memory FAQ, Anthropic’s memory guide, Google’s past chats page and Privacy Hub, Microsoft’s Copilot Memory page, and the Open WebUI docs.

Claude’s page is the most direct. “Everything Claude remembers is listed under Topics”, and each one can be opened, edited or deleted. You can also change memory by talking: tell Claude what to remember, change or forget, and the help page says the update applies to your next conversation. Memory is on by default on the Free, Pro, Max and Ultra plans.

Gemini’s memory summary is yours to edit too. Google says you can “add new details, fix inaccuracies, or delete information you no longer want saved”. Its memory of past chats has entry conditions: you must be 18 or over, signed in to a personal Google account, and keeping your chat activity. Work and school accounts do not get it.

The assistants also hold some things back. Claude’s guide says it does not store topics related to personal or sensitive subject matter by default, and that some information is never saved even if you ask. Treat that as a helpful default and still read the list yourself.

Deleting a chat does not always delete the note

This is the part most people get wrong. In ChatGPT, saved memories are stored apart from your conversations, so deleting a chat does not remove a memory that came from it. Claude behaves the same way. When you delete a conversation, “memory entries generated from it aren’t removed”, and you delete the entries one by one from the memory page.

Gemini runs in the opposite direction. Its help page says that to delete something Gemini remembered about you, you “delete all chats with this info”. It adds that there might be a short delay before Gemini stops using a deleted chat. If the detail also sits in a Google app you connected, you need to disconnect that app as well, or Gemini can find it again.

Throwing away the conversation and taking down the note are two separate jobs in most assistants. Illustration generated with Nano Banana 2.1 via Runware.

Turning memory off is a third action, and it does not clean up after itself either. Microsoft states that turning off Copilot Memory “doesn’t automatically delete already saved memories”. Claude’s pause works the same way: it keeps its existing memory but will not use it or add to it. A full wipe is a separate button. In Claude it is called Reset memory, and it permanently deletes all memories, project memories included.

Deleted does not always mean gone that second. OpenAI says it may keep a log of deleted saved memories for up to 30 days for safety and debugging. Google says a deleted chat starts being removed from its systems immediately, without giving a finish time.

For a conversation that should leave no note at all, use the mode built for it. ChatGPT’s Temporary Chat does not use existing memories or create new ones. In a Claude incognito chat, Claude “won’t remember your chats, so they won’t be saved to Claude’s memory or your chat history”. Gemini has temporary chats as well, though Google still keeps them with your account for around two months.

Memory and training are two separate switches

Memory changes what one assistant says to you. Training changes what a future model knows in general. They are different systems with different settings, and turning one off does nothing to the other.

Two switches that are easy to mix up
What it controls
Memory: Whether notes about you are saved and read back into your chats
Training: Whether your chats can be used to build the maker’s future models
Who sees the effect
Memory: Only you, in your next conversation
Training: Nobody directly. It shapes later versions for everyone.
Can you undo it?
Memory: Yes. Edit or delete the note and it is gone from the next chat.
Training: Only going forward. Turning it off stops new chats being used.

In ChatGPT the training switch is called “Improve the model for everyone”, under Data controls. While it is on, OpenAI may use your chats and remembered information to improve its models, and you can turn it off there. Anthropic says it uses consumer chats to improve its models if “you choose to allow us to use your chats and coding sessions to improve Claude”, and that incognito chats are not used even then.

Google’s setting sits under History and is named “Chat content including text & uploads”. The help page says it is on by default for adults. Gemini history is deleted automatically after 18 months unless you pick 3 months, 36 months or no auto-delete. Google also says a subset of chats is read by human reviewers, and that reviewed chats are retained for up to three years.

So an app can remember you without training on you, and train on you without remembering you. Check both. Our earlier post on what happens to your AI chats after you hit send follows the training and retention side in detail.

The notes are plain text, so they can move with you

Text is easy to carry. Anthropic’s import page gives you a prompt to paste into another assistant that asks it to write out what it knows about you. You paste the answer into Claude’s settings, and Claude stores it as individual memory entries. Anthropic labels the import experimental. Going the other way has no button. You ask Claude to “write out your memories of me verbatim” and save the reply in a file.

That prompt works anywhere, which makes it a handy backup. Official exports are patchier. Google’s data download page covers your Gemini chats, generated media, uploads and Gems, and does not mention the memory summary.

Notes that are plain text can be packed up and taken to another assistant. Illustration generated with Nano Banana 2.1 via Runware.

With local tools the notes are simply files on your own disk. Open WebUI stores memories “locally in your Open WebUI database”, scoped to your user account. Its automatic review of recent conversations is off by default. Claude Code, Anthropic’s coding tool, keeps instructions you write and notes it writes itself, and “both are loaded at the start of every conversation”. Only the first 200 lines or 25KB of its notes are read each session. On our machine they are Markdown files in a folder, and any text editor opens them.

If you build with these models, memory is your job. Anthropic’s memory tool “operates client-side: Claude requests file operations, and your application runs them”. You switch it on with one entry in the request:

"tools": [{ "type": "memory_20250818", "name": "memory" }]

Claude then asks to view, create or edit files under a /memories folder, and it checks that folder before starting a task. Your code decides where the folder lives, who can read it and when a file is deleted. A plain API call with no such tool remembers nothing between requests.

Five minutes shows you what yours has kept

Memory is useful. It saves you from repeating your job, your tools and your preferences in every chat. The cost is that a wrong or outdated note quietly bends every answer that follows, and you will not see why unless you look. Looking takes five minutes.

The five-minute memory check
1
Read the notes
Open the memory page in settings, or ask the assistant to write out what it has saved about you.
2
Fix what is wrong
Edit a stale note or tell the assistant the new fact. An old job or an old address will keep steering answers until you do.
3
Delete what you would not write down
Remove the note itself, and in Gemini the chats behind it. Then ask again to confirm it is gone.
4
Find the training switch
It lives under data or privacy settings, apart from memory. Set it the way you want.
5
Learn the no-notes chat
Temporary or incognito chats exist for the conversations you do not want remembered.

A good test for step three: would you write this on a card and hand it to a new assistant on their first day? Your role, your writing style and your current project pass easily. A passing remark about a health worry or a family argument may not, and it is yours to remove.

For conversations that should leave nothing behind, start in the no-notes mode. CSuite’s chat has one too: New temporary chat is not saved and is deleted when you leave it.

The assistant does not remember you the way a friend does. It keeps a card with a few hundred words on it and reads that card before every reply. You can ask to see the card today, and you decide what stays on it.

More reading
Black Friday sale · 25% off

One-time payment. Yours forever.

No subscriptions. No renewals. Buy CSuite once, future updates included.

Secure checkout via Stripe. Already have a license? Download the app