Context rot
Also: context degradation · context pollution · long-context decay
Quality decay that sets in when a context window fills with stale, irrelevant, or contradictory tokens.
Why CCAR-P tests this
The exam checks you recognize that piling more into the context can hurt, not help, and that curation and compaction are the fix.
Context rot is the degradation in answer quality that happens as a context window accumulates low-value, stale, or conflicting content — long meandering histories, redundant tool output, irrelevant retrieved chunks. The signal the model needs gets diluted, and it may latch onto outdated or contradictory material.
The counterintuitive lesson is that more context can make things worse. Beyond a point, adding tokens increases cost and latency while lowering precision, because the model must attend across a noisier field. This is why "just give it the whole history / whole corpus" is an anti-pattern.
Mitigations are the same disciplines as context-window management: compaction (summarize old turns), pruning stale tool results, targeted RAG instead of bulk dumps, and resetting or re-grounding a session that has drifted. Keeping the window lean and relevant preserves the model's focus.
Exam trap
"Add more context to fix a wrong answer" can backfire — if the window is already noisy, the fix is to prune and compact, not to append more.
Commonly confused with
Related concepts
Independent, unofficial study material from Claude Cert Prep. Not affiliated with Anthropic.