Context Rot in plain English.

Context rot is the observed pattern where a model's accuracy and instruction-following quality quietly degrade as its context window fills up with more tokens, even though the conversation is nowhere near the model's stated maximum length. A model advertised as supporting a million tokens does not necessarily reason about token 900,000 as reliably as it reasons about token 10,000 — instructions given early in a long conversation get followed less faithfully, relevant details buried in the middle of a huge document get missed, and answers grow vaguer the longer and more cluttered the context gets. The term captures a gap between a spec-sheet number (maximum context length) and actual usable quality at that length, which benchmark charts advertising ever-larger context windows tend not to show. The practical takeaway for anyone building with AI is that stuffing a model's entire context budget full "because you can" is often worse than curating what actually goes in — part of why context engineering has become its own discipline.