Researchers from the paper 'Why Does Claude.md Keep Growing? Catastrophic Remembering in Agentic Coding' have identified a fundamental flaw in how AI coding assistants accumulate instructions over time—and it's worse than you think. The study, posted to arXiv on August 11, 2026 by Kushal Chakrabarti, analyzed 247,694 instruction lifetimes across 1,867 real-world repositories and found that agentic prompts don't just grow—they explode without bound, tripling in size over their lifetime and gaining an average of +4.9 net instructions with every single commit.

The Memory That Never Forgets

The phenomenon has been dubbed 'catastrophic remembering'—the inverse of the well-known 'catastrophic forgetting' problem that plagues continual learning systems. While traditional machine learning struggles to retain old knowledge when acquiring new skills, agentic coding assistants face the opposite crisis: they can't stop accumulating instructions even when those instructions become obsolete or redundant. The root cause is mathematical. Appending a new instruction costs almost nothing in computational terms, but deleting an existing one requires verifying that removal won't break anything across all interactions with a prompt containing |D| instructions—resulting in a deletion cost of O(2^|D|). Translation: the larger your prompt grows, the more dangerous it becomes to prune anything from it. You're essentially trapped by your own documentation.

The Numbers Tell a Brutal Story

The data is damning. Across nearly a quarter-million instruction lifetimes studied in production repositories, prompts consistently grew by +226% over their operational lifetime—more than tripling in size. Perhaps most concerning: the older an instruction became, the less likely it was to ever be deleted. The researchers measured this via log-hazard analysis at -0.032/commit, meaning each passing commit made that instruction incrementally more entrenched and immovable. What starts as a helpful note about handling authentication becomes fossilized in your Claude.md file alongside three other contradictory notes from different team members, none of which can safely be removed without risking regressions in how the AI interprets your codebase.

Comments: The Unexpected Solution

Here's where it gets interesting for anyone actually shipping code with these tools. The researchers demonstrated that prompt comments—yes, those inline explanations developers have used since the dawn of programming—can eliminate 99.3% of excess instruction growth. By inverting IFEval to create verifiable worlds where optimal prompts are known, they showed that embedding latent reasoning directly into prompts via structured comments reduced growth from +211.3% down to just +1.4%. When applied to real-world scenarios through WildIFEval, this comment-based approach improved agentic instruction-following accuracy by up to 23.1%. The fix isn't better models or smarter context windows—it's literally teaching AI systems why instructions exist alongside what those instructions require.

Key Takeaways

  • Agentic prompts grow without bound because deletion is exponentially expensive while addition is cheap
  • 247,694 instructions across 1,867 repositories confirmed tripling +226% growth rates in production
  • Older instructions become increasingly permanent—each commit makes them harder to remove
  • Prompt comments encoding reasoning reduce excess instruction growth by 99.3%
  • Comments improved real-world task performance up to 23.1% on WildIFEval benchmarks

The Bottom Line

If English is the new code, we've been shipping it without documentation for years—and this research proves just how badly we need that to change. The fix isn't some breakthrough in transformer architecture or context management; it's comments, plain and simple. Start annotating your Claude.md files with reasoning today.