0XP
All concepts
Concept 06 of 090/3 mastered

The Context Window

The desk it works on

The model has no memory between messages. Everything it knows about your conversation is re-sent every turn, and it all has to fit in a fixed budget of tokens. Overflow the budget and the oldest material silently falls off the desk.

Everything has to fit on one desk
Page 01 / 05

Everything has to fit on one desk

The model has no memory. What it has is a working space — and everything it can use has to be inside that space at the moment you press send.

That claim needs defending straight away, because ChatGPT has a feature called Memory, Claude and Gemini have their own versions, and you may have watched one of them recall something you said last week. So either the sentence above is wrong, or something else is going on.

Something else is going on. Memory is a feature of the product, not a property of the model. When you tell ChatGPT to remember that you write in British English, the product saves that sentence to a store outside the model — an ordinary database, the kind any web app has. Later, when you start a new conversation, the product quietly pastes it back into the text it sends. The model does not recall anything. It is re-reading a note the product wrote down and handed back to it, along with everything else, as though you had typed it yourself.

The test that settles it: anything the model can use has to arrive as text, every single time. That is why memory features have their own settings page and their own list you can edit and delete from — a list is a thing a product maintains. And it is why turning Memory off does not make the model forget, it makes the product stop pasting. The working space is still the whole story; the feature just decides what gets loaded into it.

Two conversations a day apart, with a store sitting between them that belongs to neither, and watch where the second one gets its opening band from.

Two conversations, and what carries between them

Each band is one real message, as wide as the tokens it costs. Watch what the second conversation starts with.

MONDAY — THE WINDOWSYSTEMUSERASSISTANTUSERA DATABASEOUTSIDE THE MODELTUESDAY — A NEW CONVERSATIONTHE NOTEUSERNOTHING ELSE FROM MONDAY IS HERE
What this shows
Watch where the second conversation gets its first band from.

So the space holds the product's hidden instructions, anything a memory feature has decided to include, the whole conversation so far, any documents you attached, and your new message. All of it competing for the same budget, measured in tokens.

Think of it like this

Like a desk with no drawers

Everything you are working with has to be on the desk, in view. There is no filing cabinet and no drawer. When it fills up, the papers at the back slide off the edge — and nothing on the desk records that they were ever there. Somebody asks about a figure from a sheet that fell off an hour ago, and you answer from what is still in front of you.