03 · AI Fundamentals Lab
Where does a chat go when the model “forgets” it?
A model reads one fixed-size block of text every single time it answers — the standing instructions, the whole conversation so far, any documents you attached, and the room it needs to write the reply. Keep adding, and something has to come out of the back. Here you can watch exactly what goes.
Pack the window
Everything the model will read has to fit inside the bar below. Add to the conversation and watch the bar fill. Token counts are estimates from a simplified tokenizer built into this page — not a real model’s tokenizer.
These sizes are deliberately tiny so that overflow happens in a few taps. Real windows today run from a few thousand to well over a million tokens. The mechanic is identical — it just takes longer to hit.
The answer is written inside the same window. Space you fill with input is space the model cannot use to reply.
Inside the window
Outside the window — the model cannot see this
Not archived. Not summarised. Not “remembered, but hard to recall”. These lines are simply not in the text the model is handed, so for this answer they never existed.
The model has no memory. It re-reads the whole window every time, and anything pushed out simply does not exist for it.
Three ways to make it fit
Something has to go. Which something is a design decision — and every choice costs you something. Pick one and it also applies to the suitcase tab.
Same conversation, three strategies
Built from whatever you packed on the suitcase tab, at the window size you chose there. All counts are estimates.
The trade-offs
Every turn resends the entire chat
The model is not sitting there holding your conversation between messages. Each time you press send, the whole history goes back through it from the top. So turn 10 is not one message — it is ten messages, again.
My observations
Kept in this browser tab only, for this session.