Deep into one very long chat, you notice that a constraint you stated in your opening message is quietly being ignored. Knowing how Cursor handles a context window that is filling up, what is the likeliest explanation and the sensible response?

The principle — Every chat shares a fixed context window with the model, and files, tool results and messages all consume it. Cursor documents that when the window gets close to full, it compresses older parts of the conversation into a summary to leave more room for new conversation. The thread keeps working; the early detail thins out.

Why the key is correct — A constraint stated once in the first message is precisely the kind of detail a summary flattens. Nothing has failed, and no error is shown, which is why the symptom reads as the model becoming careless. The two responses that actually work are to restate the constraint now, so it re-enters the window as recent text, or to start a fresh chat for the next task so the window begins empty. This is also the argument for putting durable constraints somewhere they are re-supplied on every request rather than saying them once in a long thread.

Why the others are wrong — Treating the early turns as available verbatim leads to rewriting a perfectly clear instruction over and over, when the real problem is that its exact wording is no longer in view. Assuming the newest messages are dropped first inverts what happens and would push you towards restating recent context, which is the part still intact. Expecting a hard failure when the window fills misses the whole difficulty: the degradation is silent by design, so nothing tells you it has happened.

Remember this — Long threads forget the beginning first. Restate what matters, or start a new chat.

Sources — Cursor's agent prompting documentation.