Catastrophic Remembering: Why AI Coding Prompts Like CLAUDE.md Grow Unbounded

A new arXiv paper (2608.11095) identifies a phenomenon called 'catastrophic remembering' in agentic coding systems, where instruction files like CLAUDE.md grow without bound in real repositories. The study, based on 247,694 instruction lifetimes across 1,867 repositories, finds that prompts more than triple over their lifetime (+226%), gaining an average of 4.9 net instructions per commit. The root cause is imperfect recall: appending an instruction is cheap, but deleting one without risking correctness regressions costs O(2^|D|) in a prompt of |D| instructions. The paper also shows that prompt comments can halt growth, and inverting IFEval yields verifiable results. This work is the inverse of catastrophic forgetting in continual learning, highlighting a new challenge for AI-assisted coding.