Nib asked for
The Context Window Limit
The model only sees a fixed-size slice of text each time it answers, and it keeps no lasting memory of your conversation, so whatever falls outside that slice is simply gone for it.
5 steps, drawn by Nib. Read it here, or watch him draw it. Updated .
1the window
Recall the loop from page one: the model predicts the next token, appends it, and predicts again. Each prediction looks at the text so far. But 'so far' has a ceiling. The model can take in only a fixed number of tokens at once, and that maximum is called its context window.
The context window is the fixed amount of text the model can look at in one go.
2one budget
In a chat, the window holds everything at once: your messages, its replies, any instructions the app added behind the scenes, and the answer being written right now. All of it shares the same budget. A long reply spends room that your earlier messages needed.
Your messages, its replies, and hidden instructions all share one budget.
3sliding out
When a conversation grows past the limit, something has to go. Typically the oldest text is trimmed or cut off, so the start of your chat drops out of view. The model is not choosing to ignore it. For the next prediction, that text does not exist.
When the chat outgrows the window, the oldest text falls out of view.
a common mix-upThe model is being careless or forgetful in a human way; really the text is no longer given to it at all.
You told the chatbot your dog's name at the very start of a huge conversation, and later it asks again. What likely happened?
Show the answer
The early message fell outside the context window, so the model no longer sees it.
4the loop
Even inside the window there is no stored memory. The model's parameters, set during training, never change while you chat. The only reason it seems to remember is that the app resends the whole conversation each turn, and the model rereads it from scratch. Close the chat and start a new one, and that text is not there to resend.
It doesn't remember your chat; the whole conversation is re-read from scratch every turn.
a common mix-upThe chatbot learns from and remembers each conversation the way a person would.
If you open a brand-new chat, does the model carry anything over from your last one?
Show the answer
No. The new chat starts with an empty window, and the parameters never recorded your old conversation.
5the fix
This gives practical habits. If something matters, restate it near the point where you need it. In very long chats, paste key facts or a short summary into a fresh conversation. If answers drift or contradict earlier details, suspect the window before suspecting stubbornness.
Restate what matters, and start fresh with a summary when chats get long.
The model isn't remembering you; it's rereading a limited stretch of text each time, so what you want it to use must be inside that window.