# Raw Provenance and Distilled Guidance Serve Different Memory Jobs ## Claim A knowledge base only stays useful when something cheap sits in front of it deciding what to retrieve, so raw provenance and distilled guidance need to be kept as separate artifacts with separate lifecycles rather than merged into one growing file. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds when retrieval is cheap and exact and the index stays small enough to always load — true for project-scale file corpora. - Breaks when the corpus outgrows a flat index and needs hierarchy or embeddings. - Breaks when descriptions are too vague to route on, or when entries go stale and the router confidently sends the reader somewhere wrong — nothing in the base pattern prunes stale entries. ## Evidence A coding-agent workflow logs every exchange with the commit hash of the resulting change (raw provenance, append-only, auditable), then periodically has the agent distill generalized lessons into a separate field-guide directory (distilled guidance, meant to stay small and current), fronted by an `init.md` index of links and one-line descriptions so the agent retrieves only relevant entries instead of loading the whole corpus into context. ## Links - [[3. Resources/Permanent Notes/Capable Models Need Fewer Constraints Not More]] — because extends: the same underlying shift — capable agents need less upfront loading and more on-demand retrieval — from a different source in the same week - [[3. Resources/Permanent Notes/Intent Written on the Outside Compounds While Re-Derived Intent Is Debt]] — because applies-in: the field guide is one concrete implementation of externalizing intent so it isn't re-derived every session - [[3. Resources/MOCs/Personal Knowledge Management MOC]] — because applies-in: this vault's own MOCs are structurally the same pattern — an index of links and short descriptions in front of a larger note corpus ## Tests / falsifiers - A/B the loading strategy — full preload vs. `init.md` index plus tool-call retrieval — measuring output quality, total tokens consumed, and latency; expect a quality/cost frontier rather than a clean win either direction. - Longitudinal test: sample guide entries periodically and label each still-accurate / stale / actively wrong; correlate stale-entry proportion with downstream error rate. This is the decay question the base pattern doesn't answer. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/The Field Guide Technique — Persisting Agent Feedback Across Sessions]] — the four-stage technique and its `init.md` router.