Each time the agent calls the model, it sends a single bundle of text: the context window. The model can use only what is in that bundle. It has no hidden notebook and recalls nothing from the previous call.
The window has a fixed size, counted in tokens (roughly word pieces). We will use an illustrative 8,000-token window. Our agent is a trip planner with the goal "Plan 3 days in Lisbon under 900 euros", and its instructions plus goal take 1,000 tokens.
💡Good to know
Real windows are larger, but the same arithmetic applies. Only the numbers change.