Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 12·5 min read
Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding
TL;DR
Agentic coding READMEs, like CLAUDE.md, tend to grow indefinitely due to a phenomenon called catastrophic remembering, where deleting old instructions becomes increasingly difficult. By introducing prompt comments, researchers found they could significantly reduce unnecessary instructions and improve instruction-following performance.
✦ Why It Matters
Start using prompt comments in your agentic coding projects to enhance clarity and reduce instruction bloat.
Key Takeaways
- Implement prompt comments to clarify instruction rationale.
- Monitor instruction growth in agentic coding projects.
- Consider the impact of instruction age on deletion likelihood.
- Explore methods to reduce unnecessary instructions in READMEs.
How It Works
The concept of catastrophic remembering arises from the difficulty of deleting instructions once their context is lost. As instructions accumulate, the cost of removing them increases exponentially, making it easier to add new instructions instead.
By using prompt comments, which provide context and reasoning, the need for excessive instructions diminishes, allowing for clearer and more concise coding.