Why CLAUDE.md Keeps Growing? Paper Reveals 'Catastrophic Remembering' in Agents
rohanpaul_ai · x · 2026-08-14
A study analyzing 1,867 GitHub repositories highlights a ratchet problem in agentic instruction files (like CLAUDE.md): instructions are easily added but hard to delete.
- The Phenomenon: Instruction files tripled in size on average (growing 226%), with older instructions becoming less likely to be removed. The paper coins this "catastrophic remembering," where instructions survive long after their original reasoning is forgotten.
- The Fix: Researchers propose a classic software engineering practice: attach comments documenting the failure, hypothesis, and outcome behind each instruction, then hide these comments from the model.
- Results: In a 51-step controlled eval, informative comments slashed excess prompt size growth from +211.3% down to +1.4% without degrading constraint satisfaction.
Related event: Study Reveals 'Ratchet Effect' Causing Infinite Growth in CLAUDE.md(2 posts)→
More from coding & agent
- Terminal Night 3 talk slides released, missing demo video to follow — 4310sy · 2026-10-16
- Agent developers are sitting on unused API credits that can exceed actual usage — EnvironmentalDay8864 · 2026-10-16
- Notion's bet on agents: abstract away deployment so users and agents just write code — simonguozirui · 2026-10-16
- Four Gemini 4 Argon agents beat Minecraft's Ender Dragon autonomously in 27 hours — Maxence Faldor · 2026-10-16
- Tencent's Terminal-RSI builds recursively self-improving terminal agents, hitting 19.84% on Terminal-Bench 4.0 — tencent · 2026-10-16
- Tsinghua's AutoSFT uses multi-agent collaboration to automate SFT data research — TsinghuaNLP · 2026-10-16