Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding

Source

arXiv

arxiv.org/abs/2608.11095v1

Read the source

Source

  • Source name: arXiv
  • Source kind: paper
  • Authors / organization: Kushal Chakrabarti
  • Venue: arXiv: cs.AI
  • Published: 2026-08-11
  • Collected: 2026-08-12
  • URL: http://arxiv.org/abs/2608.11095v1
  • PDF: https://arxiv.org/pdf/2608.11095v1
  • DOI: UNKNOWN
  • Registry group: academic_apis_and_preprints
  • Evidence role: canonical_academic_index
  • Language: en
  • Country / region: UNKNOWN

Plain-Language Brief

The research notes discuss a phenomenon called 'catastrophic remembering' in Agentic coding READMEs like CLAUDE.md where instructions grow without bound and do not get deleted easily over time. The paper proposes that adding comments to the prompts can help reduce this growth and improve the performance of agentic instruction-following by up to 23.1%.

What the Document Says

  • The text discusses the issue of Agentic coding READMEs like CLAUDE.md growing without bound over time, with the growth being more pronounced in older instructions.
  • It is stated that this phenomenon is due to imperfect recall: appending an instruction is cheap but deleting one without risking a correctness regression costs O(2^|D|) in a prompt of |D| instructions.
  • The term 'catastrophic remembering' is introduced as the inverse of catastrophic forgetting, which is the basis for continual learning.
  • The paper reports findings from an analysis of 1,867 repositories and 247,694 instruction lifetimes, stating that agentic prompts grow without bound, with older instructions less likely to be deleted (log-hazard -0.032/commit).
  • The text mentions that prompt comments can halt the growth of these prompts and remove 99.3% of excess instructions.
  • It is stated that inverting IFEval yields verifiable worlds whose optimal prompts are known, and that comments encoding latent reasoning can improve real-world agentic instruction-following by up to 23.1%.

Key Concepts Mentioned

  • catastrophic remembering
  • Agentic coding
  • CLAUDE.md
  • imperfect recall
  • prompt comments

Claims Made by the Document

  • Agentic prompts grow without bound over their lifetime.
  • Older instructions are less likely to be deleted than newer ones.
  • Prompt comments can halt the growth of agentic prompts and remove 99.3% of excess instructions.
  • Inverting IFEval yields verifiable worlds.
  • Comments encoding latent reasoning can improve real-world agentic instruction-following by up to 23.1%.

Evidence Used

  • Analysis of 1,867 repositories and 247,694 instruction lifetimes

Methods / Approach

  • Data analysis
  • Inversion of IFEval

Limitations Stated by the Authors

  • The study focuses on Agentic coding READMEs like CLAUDE.md and may not be applicable to all types of code or programming languages.

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • The text does not provide information on the specific implementation or details of the IFEval method.
  • The study does not explore potential drawbacks or unintended consequences of using prompt comments.

Open Questions

  • What are the potential drawbacks or unintended consequences of using prompt comments?
  • How can the findings from this study be applied to other types of code or programming languages?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • It is important to note that while the text mentions that 'English is the new code', it does not elaborate on what this means in the context of the study.

Raw Abstract / Excerpt

Agentic coding READMEs like CLAUDE.md grow without bound in real repositories, stopping only when the repository retires or someone rewrites the file wholesale. We trace this to imperfect recall: appending an instruction is always cheap, but once an instruction's rationale is gone, deleting it without risking a correctness regression costs O(2^|D|) in a prompt of |D| instructions. We name the resulting divergence catastrophic remembering, the inverse of catastrophic forgetting around which continual learning is organized. First, we characterize this phenomenon across 247,694 instruction lifetimes in 1,867 repositories: agentic prompts grow without bound, more than tripling over their lifetime (+226%), gaining +4.9 net instructions every commit; further, the older an instruction gets, the less likely it is to be deleted (log-hazard -0.032/commit). Then, we show that prompt comments can halt the growth: inverting IFEval yields verifiable worlds whose optimal prompts are known, and there comments encoding latent reasoning remove 99.3% of excess instructions (+211.3% to +1.4%). Finally, applying the same inversion to WildIFEval, we show that prompt comments can improve real-world agentic instruction-following by up to 23.1%. If English is the new code, why don't we have comments yet?