# 论文提出"灾难性记忆"概念：CLAUDE.md 等智能体指令文件为何持续膨胀

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-07 23:01
- AIHOT 分数：61
- AIHOT 链接：https://aihot.virxact.com/items/cmtrdodf5000nro8dej2qwd31
- 原文链接：https://x.com/rohanpaul_ai/status/2096977141326139447

## AI 摘要

一项跨 1,867 个 GitHub 仓库的研究发现，CLAUDE.md 类智能体指令文件在生命周期内平均增长 226%，且指令越老越难被删除，作者称之为"灾难性记忆"，即指令留存而当初的理由消失。论文提出为每条指令附加记录失败、假设与结果的注释并对执行模型隐藏，在 51 步 IFEval 设置中将多余提示词规模从 +211.3% 降到 +1.4%，注释形噪声无法复现该效果。

## 正文

New term “catastrophic remembering”

CLAUDE.md has a ratchet problem: instructions are easy to add and increasingly hard to delete.

Agent prompts may remember rules much longer than maintainers remember why those rules exist.

Across 1,867 GitHub repositories, agentic instruction files more than tripled over their lifetime, with instruction count rising 226% on average.

Deletion became less likely as instructions aged, and that decay was steeper in files touched by multiple human authors.

The paper calls this “catastrophic remembering”: the instruction survives, but the reasoning that once justified it disappears.

The proposed fix is a decades-old software practice: attach a comment recording the failure, hypothesis, and outcome behind each instruction, then hide that comment from the executing model.

In the 51-step controlled IFEval setting, informative comments cut excess prompt size from +211.3% to +1.4%, with last-three-round constraint satisfaction at 44.0% in both arms.

Comment-shaped noise did not reproduce the effect, pointing to preserved rationale rather than extra text.