Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
Data Science
Towards Data Science·22d ago·5 min read
Context Rot: Why Claude Code Sessions Decay, and How to Govern Them
TL;DR
Context rot occurs when the quality of outputs from AI models like Claude Code declines over time due to the degradation of the context window. Instead of simply increasing context, the article suggests implementing governed context to maintain session quality.
✦ Why It Matters
Implement governed context in your AI sessions to prevent context rot and enhance output quality immediately.
Key Takeaways
- Curate context inputs before starting a session.
- Reset the session if confusion persists after multiple corrections.
- Keep context minimal and high signal to avoid dilution.
- Use external files for durable state to prevent stale information.
How It Works
Claude Code's context window is a sequence of tokens that includes prompts, responses, and tool calls. Each token competes for attention, which dilutes the model's ability to focus on relevant information as the context grows.
This dilution leads to a gradual decline in output quality, especially in the middle of long sessions where retrieval accuracy is lowest.