Because adding a quick instruction to an AI’s guide file is fast and low-risk, developers continually tack on new rules whenever the AI makes a mistake (e.g., “Don’t use library X,” or “Always run tests with flag Y”). Over time, however, the original context and reasoning behind older instructions fade from human memory.

Agent instruction files like CLAUDE.md exhibit unbounded growth due to “catastrophic remembering,” a process where the original rationale for instructions is lost, making safe deletion difficult…