'Catastrophic remembering': CLAUDE.md files grow +226% over their lifetime and almost never shrink — and prompt comments fix it

arXiv·high signalXBlueskyLinkedInCopy link

Analysis of 247,694 instruction lifetimes across 1,867 repositories finds agentic instruction files grow without bound, more than tripling over their lifetime and gaining +4.9 net instructions per commit, with older instructions progressively less likely to ever be deleted (log-hazard -0.032/commit). The authors name this catastrophic remembering — the inverse of catastrophic forgetting — and trace it to a cost asymmetry: appending is cheap, but deleting an instruction whose rationale is lost costs O(2^|D|) to verify safely. Adding comments that encode the latent reasoning behind each instruction removed 99.3% of excess instructions in verifiable settings and improved real-world instruction-following on WildIFEval by up to 23.1%.