WTF is Context Rot?
Models get measurably worse as their context grows, even when the extra tokens are irrelevant.
Also called: context degradation, long-context degradation
TL;DR
A 1M-token context window doesn't mean the model uses all of it equally well. As you stuff more into the prompt, accuracy quietly drops, the model gets distracted by old history, and it latches on to the wrong details. This is why long agent sessions start to feel "dumber" over time.
How it works
Chroma's report tested 18 models (including GPT-4.1, Claude 4, Gemini 2.5, Qwen3) and found performance degrades non-uniformly with input length, even on trivially simple tasks such as repeating words; semantic (rather than lexical) needle-in-a-haystack retrieval and the presence of distractors made it worse. Drew Breunig's earlier taxonomy names four failure modes: context poisoning (an error enters the context and keeps being referenced), context distraction (the model over-relies on history instead of what it learned in training), context confusion (irrelevant content, such as too many tool definitions, degrades answers) and context clash (accumulated information contradicts itself). The practical response is context engineering: compaction, just-in-time retrieval, fewer tools, and fresh contexts for new tasks.
Origin
- Coined / introduced by
- Chroma (Kelly Hong, Anton Troynikov, Jeff Huber); failure modes by Drew Breunig
- When
- 14 Jul 2025
- Story
- Chroma's 'Context Rot' report gave the phenomenon its name; Breunig's 'How Long Contexts Fail' (22 Jun 2025) catalogued the failure modes.