Context Rot

Context rot is the gradual decline in a language model's accuracy and instruction-following as its input grows longer, even when that input stays well inside the advertised context window. The causes are still being studied, but likely contributors are attention spread over more tokens, models seeing fewer long sequences in training, and distractors or stale earlier turns competing with the facts that matter. A context window limit is a hard cutoff, whereas context rot is a quality loss that starts long before the limit is reached. In agent systems it shows up in long sessions where tool outputs, logs and old plans pile up, and it is one of the reasons, alongside the hard window limit, that agent harnesses compact, summarize or offload context.

The term was popularized by a 2025 Chroma technical report that measured the effect across current models on deliberately simple retrieval and repetition tasks. Common mitigations are keeping only what the next step needs, replacing verbose tool output with short summaries, restating key instructions near the end of the prompt, and moving long-lived facts into retrieval or memory instead of the running transcript.

Work with us

Ready to put agentic AI to work?

Book a free 30-minute consultation. We'll map one real process worth automating with production-grade AI.