Context Rot

Definition

The degradation of accuracy and recall as the token count in a context window grows — "As token count grows, accuracy and recall degrade, a phenomenon known as context rot."

Key points

• Curating what is in context is just as important as how much space is available.
• Applies even on 1M-window models — a large window does not prevent rot.
• Motivates context hygiene: Server-side Compaction, the Scratchpad Pattern, structured summaries, and isolated subagent contexts.
• Isolated subagent context prevents context rot and enables parallelism (each subagent gets a clean window) — see Subagent decision criteria.

Why it matters for the exam

• The conceptual "why" behind curating context, spawning subagents, compaction, and scratchpads. Expect it as the justification in long-session and multi-agent scenarios.

Common gotchas

• "Just use the 1M window" is the wrong answer — more tokens can make recall worse, not better.
• Distinct from Lost In The Middle: context rot is about total volume; lost-in-the-middle is about token position.

See also

• Context Engineering
• Lost In The Middle
• Server-side Compaction
• Scratchpad Pattern
• Context Windows By Model
• Subagent decision criteria

Sources

• docs-models-pricing
• netlight-study-guide
• blog-context-engineering