AI agent glossary
Context rot
The tendency of a model's accuracy and instruction-following to degrade as its context fills with more tokens, even well below the window limit.
A large context window does not mean every token gets equal attention. As the input grows longer, models tend to become worse at finding and using specific details, following earlier instructions, and ignoring irrelevant material. The term context rot became popular in 2025 to describe this gradual decline.
Agents are especially exposed because their context grows with every step: tool outputs, file contents, failed attempts, and long conversation history. A long session can drift, repeat mistakes, or forget rules stated at the start.
The remedies are the core of context engineering: keep tool outputs small, compact or summarize history, start fresh sessions for new tasks, move durable facts into notes or memory files, and delegate side investigations to subagents so their clutter never reaches the main context.
Learn more in the course
Related terms
Go from definitions to doing
Learn to use and build AI agents in 65 practical lessons.