Context Rot

The progressive degradation of Claude Code’s output quality as session length grows, caused by attention dilution across an ever-larger token history.

Key points

  • Retrieval accuracy drops from 92% at 256K tokens to 78% at 1M tokens [011]
  • Analysis of 18K thinking blocks across 7K sessions: thinking depth dropped 67% and “edit without reading” behaviour rose from 6% to 34% in long sessions [011]
  • One developer tracked 100+ message chats and found 98.5% of tokens were spent re-reading old conversation history — not producing new output [011]
  • Token cost is exponential: message 30 in a chat can cost 31x more tokens than message 1, because the entire history is re-read each turn [011]
  • Auto-compaction fires at 95% context use and retains only 20–30% of detail — the worst possible moment (most degraded state) [011]

Symptoms

  • Model edits files without reading them first
  • Contradicts earlier decisions made in the same session
  • Vague or repetitive outputs
  • Increasing error rate on tasks it handled correctly earlier in the session

Remediation patterns

  • Cap Opus sessions at ~120K tokens (12% of 1M window), then summarise and /clear [011]
  • Use /session-handoff skill to generate a structured context handoff document before clearing [011]
  • Use /re (rewind) to drop failed attempts that silently pollute future responses [011]
  • Convert documents to markdown before ingestion (33–90% token reduction) [011]
  • Use /btw for side questions that don’t need to enter conversation history [011]
  • Session chaining: discovery → planning → execution sessions, each consuming only the prior session’s output document [011]

Related entities

  • Claude Code — the primary environment where context rot occurs

Related concepts

Source references

  • [011] Nate Herk — Claude Code power features cluster (2026-04-20 to 2026-04-27)