Paper: AI coding prompts grow 226% due to 'catastrophic remembering'; comments can cut excess by 99.3%
rohanpaul_ai · x · 2026-09-07
The arXiv paper 'Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding' by Kushal Chakrabarti investigates the unbounded growth of instruction files (e.g., CLAUDE.md) in AI coding. Analyzing 247,694 instruction lifetimes across 1,867 repositories, the author finds that agentic prompts grow without bound, more than tripling (+226%) and gaining +4.9 net instructions per commit; older instructions are less likely to be deleted (log-hazard -0.032/commit). The phenomenon is termed 'catastrophic remembering,' the inverse of catastrophic forgetting. The proposed fix is to attach comments encoding latent reasoning, hidden from the executing model. In controlled IFEval tests, comments removed 99.3% of excess instructions (+211.3% to +1.4%); on WildIFEval, comments improved real-world instruction-following by up to 23.1%. The paper asks: 'If English is the new code, why don't we have comments yet?'
More from coding & agent
- He benchmarks new models by having them build their own copilot — $28 of compute — bucolucas · 2026-10-27
- Should agents be allowed to change plans mid-task? Reddit debates the drift boundary — wistfulcracker84 · 2026-10-27
- HeyGen + Canva MCP turns slide speaker notes into an avatar presentation video — HeyGen · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- Cloudflare Computer adds agent interruption recovery and approval pauses in v0.7.0 — irvinebroque · 2026-10-27
- LLM Tails cuts agent input tokens 80% by making agents read before they write — savadks1818 · 2026-10-27