# A Durable-Context File Needs Active Pruning Not Just Careful Writing ## Claim A project-memory file (CLAUDE.md, a team wiki, an onboarding doc) degrades in usefulness as the underlying project changes, so its value depends on ongoing pruning against the current codebase, not on how well it was written the first time. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Applies to any always-loaded context file meant to persist across many sessions or readers, where the described reality (commands, architecture, conventions) can drift out of sync with the file's claims. - The failure mode isn't just "missing new information" — a stale entry can actively contradict current reality and mislead a reader or agent who trusts it. - Doesn't specify a review cadence — the source asserts the decay dynamic without proposing a maintenance schedule. ## Evidence Oleg Efimov argues CLAUDE.md should hold only durable, non-obvious operational context (commands, architectural exceptions, recurring gotchas), with the explicit warning that "a static CLAUDE.md decays" — implying the sorting exercise (what earns a line) has to be repeated as the repository evolves, not performed once at file creation. ## Links - [[3. Resources/Permanent Notes/Capable Models Need Fewer Constraints Not More]] — because contrasts; that note argues for removing rules as model capability grows, this note argues for removing rules as project reality changes — two different reasons to prune the same kind of file. - [[3. Resources/Permanent Notes/Intent Written on the Outside Compounds While Re-Derived Intent Is Debt]] — because qualifies; written intent only compounds if it stays accurate — an unpruned, decaying file is exactly the case where "written on the outside" stops compounding and starts misleading. - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — because applies-in; a concrete maintenance practice for any persistent-memory file used by an AI coding agent. - [[3. Resources/Permanent Notes/A Shared Routing Budget Makes Each Added Option Degrade the Rest]] — the mechanism that makes pruning a selection-quality issue rather than a tidiness one. ## Tests / falsifiers - Audit a CLAUDE.md (or equivalent) that hasn't been reviewed in 6+ months against the current codebase and count contradicted vs. still-accurate lines — a high contradiction rate would support the decay claim directly. - Find a long-lived project-memory file that stayed accurate with zero active maintenance (e.g., because the underlying project itself was frozen) — would bound the claim to actively-changing projects specifically. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/Navigating Claude Code - CLAUDE.md Done Right]] — states the decay claim directly and frames the sorting exercise as ongoing.