What a context window actually limits

A context window is a budget, not a hard drive. The model sees a finite string of tokens for this request. It does not keep a private diary of the ones you left out.

Tokens, not words

Providers count tokens, pieces of text that may be a word, part of a word, or punctuation. English often lands near four characters per token. Code and other languages can be denser. The window is a token cap, so the same page of words can cost more in one language than another.

The cap includes everything in the request: the system prompt, the user message, attached files, and, on a chat, the earlier turns you sent back.

What falls off

If the prompt is longer than the window, the extra tokens are not in the model’s input. Typical APIs reject the request or drop the oldest turns. Either way, the model cannot “remember” text it never received.

That is why a long chat suddenly loses an instruction from the first message. The first message is no longer inside the window. It was not stored on the server as a fact. It was only in the payload you used to send.

What it is not

A larger window is not a larger trained model. It does not mean the model read the public web last night. It means this one request can carry more tokens before something is cut.