Picture a desk with room for a fixed number of pages. Everything the model needs in order to answer has to be lying on that desk at the moment it answers: the standing instructions it was given, whatever was retrieved on your behalf, every earlier turn of the conversation, and your actual question. That span is the context window, and it is measured in tokens rather than words. It is not storage and it is not memory. It is the whole of what exists for the model in the instant it replies.
The word window invites you to imagine looking out at something larger, and there is nothing larger. Whatever is not on the desk has not been forgotten — it was never there. This produces a failure with no symptom. When a conversation grows past the limit the earliest turns fall away, and the model does not announce it, cannot notice it, and will answer with exactly the composure it had while it could still see them. You asked it to hold to a constraint forty messages ago; it is not refusing you and it is not lying. The constraint is simply outside the world.
Because the window is finite and everything competes for the same room, its size has become a headline figure — a hundred thousand tokens, a million, and climbing. Two things are worth knowing about those numbers. The first is the conversion: a token runs to about three-quarters of a word in English, so a hundred thousand of them is roughly a long novel. The second is that capacity and use are different questions. A model given an enormous window still draws less reliably on material buried in the middle of it than on material near either end, and every token in the window is paid for again on every turn. Which is why deciding what goes in has a discipline of its own, and why the answer to a full window is rarely a bigger one.
The human version is the difference between what you know and what you happen to be holding. A colleague with thirty years in the field, handed the wrong three pages and asked in a hurry, will give you a worse answer than a novice handed the right one. Their knowledge did not shrink; their view did. The window is that view, drawn with a hard edge — and the uncomfortable part is that from the inside, a narrow view and a wide one look exactly the same.
