Context Collapse
The point where an AI session gets too long and the model starts mixing up, forgetting, or contradicting what you told it earlier. It also goes by "context rot", the term Anthropic uses in its context engineering guide: the more tokens fill the window, the less accurately the model recalls what it holds. It is not a bug: the context window has filled with noise (old messages, whole files, back-and-forth), and the useful signal drowns in it. You spot it when the agent re-asks questions already settled, reintroduces a bug you just fixed, or loses track of the original goal.
Strengths
- Explains in one sentence why an agent "gets dumb" on long sessions
- The fix is free: starting clean costs 0 extra tokens
Limitations
- Hard to measure precisely: you feel it before you can prove it
Best for
- Knowing when to cut an agent session and restart from a summary