How It Works
Context rot
The tendency of a language model's performance to degrade as its input context grows longer. Even on simple tasks, such as finding a fact in a long document or repeating text exactly, models can become less accurate as more tokens fill the context window, and the decline is often uneven rather than steady. It is one reason long agent conversations and very large pasted documents can produce worse answers than short, focused ones.
Origin · no single documented coiner
The phrase circulated informally among AI practitioners in mid-2025, with no single documented coiner. Chroma's July 2025 research report 'Context Rot: How Increasing Input Tokens Impacts LLM Performance' gave the term wider currency by measuring the effect across a range of models and tasks.