AI
Context Rot: Why Claude Code Sessions Decay, and How to Govern Them
In an analysis of agentic AI coding, Jake Minns details the phenomenon of "context rot" in Claude Code, where long sessions experience a gradual decline
Key takeaways
- Claude Code sessions suffer from context rot, a gradual decline in output quality that occurs well before token limits are reached.
- Model retrieval accuracy across a long context window traces a U-curve, meaning information in the broad middle is least reliably recalled.
- Models weight context heavily and will reliably incorporate irrelevant distractors, which measurably drops reasoning performance.
- The fix for session decay is not larger context windows, but governed context using commands like clear, compact, rewind, and subagents.
In an analysis of agentic AI coding, Jake Minns details the phenomenon of "context rot" in Claude Code, where long sessions experience a gradual decline in output quality well before reaching official token limits. This decay is driven by intrinsic rot (imperfect model attention and retrieval accuracy drops in the middle of long contexts) and content rot (the accumulation of stale, wrong, or contradictory information). To combat this, developers must actively govern context by limiting tool scopes, using planning modes, pushing verbose tasks to subagents, and starting fresh sessions with distilled handoff briefs.
In their words
“The fix isn't more context, it's governed context.”
By the numbers
- 100k
- Token count where a model can find a fact but struggles to reason with it
How it unfolded
- Jake Minns publishes analysis on Claude Code context rot.
Turn stories like this into views
Ravenclip finds the AI news, makes the video, and posts it before attention moves on.
Common questions
- What happened with Claude Code?
- Claude Code sessions suffer from context rot, a gradual decline in output quality that occurs well before token limits are reached.
- Where can I read the original report?
- Read the full report at towards_data_science.
