07 Lesson 7 of 15
Three kinds of memory
“Memory” is one word for three different things that last for different lengths of time. Most confusion about privacy and memory comes from mixing them up.
It remembers me, so it must be learning from me. Last week you told it you are vegetarian, and today it suggested a vegetarian recipe without being asked. So it must be learning from your chats, and what you typed must now be part of it. That is either comforting or alarming, depending on what you typed.
The model reads them like anything else you typed. It has not changed at all.
See it in the lessonIt may well remember you. But the model is not what is doing the remembering. There are three separate stores, and they are worth naming, because most everyday questions (will it remember this, is my data in it, can it look that up) become easy once you know which one you are asking about.
| Store | What is in it | How long it lasts |
|---|---|---|
| What it learned in training | Patterns from a very large amount of text, gathered up to some cut-off date. | Fixed. Set before you arrived. Talking to it does not change this. |
| This conversation | The recent turns, plus any instructions the tool adds that you cannot see. | Until it falls out of the window, or you start a new chat. |
| Things the app brings in | Saved memories about you, earlier chats it can search, web results, files you upload. | As long as the app keeps them. Saved memory lasts until you delete it, and on most major chatbots it is now on by default. |
Here is the same diagram again, now showing the whole picture. The large shell never changes as you talk. The bright box is the conversation. The separate object to the side is everything the app can bring in. Switch it on, then start a new chat: a saved note about you is copied into the fresh conversation before you type a word.
Read this diagram as text
Three controls. Add a turn puts a message into the conversation store. Start a new chat empties that store and changes nothing else. Switch on memory and search connects the third store. While it is on, each new chat starts with a saved note about you (no oven, likes quick meals) already copied into the conversation. The line under the buttons reports all of this. Nothing you do here ever changes the training store.
So is it learning from you?
No, but it can look that way. Within one chat, everything you said sits in the context window and is read again each turn. Across chats, most major chatbots now keep saved memories about you, and some can search your earlier chats too. When you start a new chat, the app copies the relevant bits into the conversation. The model reads them like anything else you typed. It has not changed at all. Delete a saved memory, and it is gone from every future chat.
This matters in two ways. You can usually open these memories, edit them, delete them, or turn memory off, often in the settings under personalisation or memory. Many tools also offer a temporary or private chat that skips memory entirely. And separately, the company may store your conversations and use them to train future versions. That is a business policy, not how the model works, and it differs between products and plans. The lesson on your data comes back to this.
Start a new chat and ask it: what do you remember about me? Then find the memory screen in your tool's settings and read what it has saved. Delete anything you would not want carried into every future chat.
Worth flagging: This lesson rests more on how these systems are built than on studies of how people misunderstand them. Memory features also change often: what each chatbot saves, whether it can search past chats, and whether memory is on by default all vary by product, plan and country. Check the settings of the tool you use.