What is context rot?

Context rot is when an AI model's answers get less reliable the more you pack into its context window, even though everything you added technically still fits under the stated limit.

It's a different problem from the context window itself. The context window is about capacity — how many tokens a model can hold at once. Context rot is about quality — what happens to the model's attention once that space starts filling up.

Researchers have observed that models get worse at using information buried in the middle of a long input, tend to lose track of instructions given earlier in a long chat, and get thrown off by piles of now-irrelevant back-and-forth that piled up along the way.

Why it matters

A bigger context window sounds like a straightforward upgrade — more room, more capability. But fitting and working well are not the same thing. Stuffing a model's context with every document, every past message, and every side note can make its answers worse, not better, because the details that actually matter get buried in the noise.

The practical fix is to treat context like a workspace you tidy, not a warehouse you fill. Trim what's no longer relevant, summarize long conversation history instead of keeping the full transcript, and start a fresh conversation once an old one has drifted far from the task at hand. A shorter, cleaner context often gives a more reliable answer than a longer one stuffed with everything you've ever discussed.