Context rot
The tendency of a model to attend less to material buried in the middle of a long context than to its start or end.
Also written: lost in the middle.
A million-token context window (The total tokens a model can consider at once: your prompt, the conversation so far, and the answer it is writing.) does not mean a million tokens of equal attention. In practice, models reliably use what is at the very beginning and the very end of their context and are measurably worse at material buried in the middle.
The practical consequences are simple. Put the instruction that matters at the end, not halfway up. Put reference material the model must not miss at the start. Treat a long context as a filing cabinet that is easy to put things into and hard to find things in, and prune it rather than letting it accumulate.