The context window, measured in tokens, tells you how much fits into a single request. Whatever does not fit, the model does not "forget" — it never sees it at all. That is what explains long conversations going dim: the earliest parts drop out, and the model answers as though they had never been said.
It has two practical consequences. One is the bill: you pay for input tokens as well, so every unnecessary attachment costs money. The other is language — tokenisers were built mainly on English, so the same document in any other language takes up more room and costs more.
The most common misconception is that the advertised window size is the usable window size. In a measurement across twelve models, where the question and the sentence being looked for shared no literal wording, at 32 thousand tokens ten of the twelve fell below HALF of their own short-context result — in one case from 99.3% to 69.7%.
How tokens and the window fit together, and what all of it does to your bill, is what Tokens and the Context Window: Why Length Costs Money sets out. Directly related terms: token, context engineering and RAG.