Context window
How much text a model can hold in view at once, counted in tokens: the conversation, the files, the tool results, and the answer being written.
Everything shares one budget. A long file read at the start is still there at the end, taking room from the answer, which is why a session that started fast can slow down and grow expensive without anything new happening.
Bigger is not automatically better. Recall drops in the middle of very long inputs, and every token in the window is paid for on every turn. Feeding a model only what the task needs beats feeding it everything you have.
When the window fills, the choices are to summarize the old part, to start clean, or to move the material out of the prompt and behind a tool that fetches it on demand.