Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding
The paper argues that agent coding instruction files bloat because old instructions become too risky to delete once their reasons are forgotten.
Across 1,867 repositories, the author reports agentic prompts more than tripled over their lifetimes, adding a net 4.9 instructions per commit. Older instructions were less likely to be removed, which the paper calls “catastrophic remembering.” In controlled prompt tests, comments that preserved the reason for instructions cut excess instructions from +211.3% to +1.4%. The same approach improved real-world instruction-following by up to 23.1%. ArXiv · AI/CL/LG's note
Across 1,867 repositories, the author reports agentic prompts more than tripled over their lifetimes, adding a net 4.9 instructions per commit. Older instructions were less likely to be removed, which the paper calls “catastrophic remembering.” In controlled prompt tests, comments that preserved the reason for instructions cut excess instructions from +211.3% to +1.4%. The same approach improved real-world instruction-following by up to 23.1%. ArXiv · AI/CL/LG's note
score 5