What is it?
Context rot
The observation that a model's quality drops as its context window fills up, even when the total tokens are well within the technical limit. In other words, the "effective" context window where the model still performs well is usually much smaller than the advertised limit. A model rated for 1 million tokens may start degrading well before that. This is why agents actively manage what stays in context rather than just letting it grow.
From the AI agent glossary
Where Context rot is explained
- How Context Goes Wrong: Rot, Pollution, and ConfusionContext Engineering · Context rot: the most important failure to understand
- What Context Engineering Is, and Why It Replaced Prompt EngineeringContext Engineering · Why this matters more in the agent era
- The Four Types of MemoryAgentic Memory · Working memory
- Governing Evolving Memory (the SSGM Framework)Agentic Memory · Key takeaways
- Single Agent versus Multi-Agent: The Core TradeoffMulti-Agent Systems · What a single agent gives you