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
- Building a neutral Codex-class harness requires running the original inner harness, dev says — joshalbrecht · 2026-09-07
- Dev Builds Timestamped Summary Site for All MLOps Community Talks, With an MCP Server — strickvl · 2026-09-07
- Shipping an Obsidian plugin with GPT-6 Astra: scan limits, SVG icons, review delays — vista8 · 2026-09-07
- Grok Build 1.0.22 ships subagent messaging, first-party MCP server and safer Auto mode — mark_k · 2026-09-07
- Agents on a 16GB MacBook Air M5: Qwen 3.5-9B 4-bit can't yet build a working app — Fluid-Author-9566 · 2026-09-07
- Remastering old i2v videos with MiniMax H3: a depth-map + reference workflow — WaitAcademic1669 · 2026-09-07