Promotion-based knowledge curation

The shape is a graduation pipeline. The bottom tier is local and unreviewed — nothing leaves the machine, nothing is gated, capture is cheap. The middle tier is a shared team store fronted by a human review queue. The top tier is a public commons that only nominated, generalized units reach. The generalization step is explicit rather than incidental: an instruction telling agents to strip project specifics as a unit moves up, so the thing that graduates is a pattern rather than one team's incident. The same instinct shows up without any software behind it — one practitioner promotes only what proved important and later distils roughly ten to fifteen dense team wikis, with about 80% of the raw material never leaving him.

Promotion handles what goes in; a confidence loop handles what stays. A recalled unit carries a typed schema — domain, insight, action taken, metadata, an optional pattern — and a confidence score. The accompanying discipline tells agents to query before starting work in a domain, validate before acting, confirm on success so the score accumulates, and flag anything stale or wrong. Shared memory then ages by evidence rather than being silently trusted forever.

The failure this avoids is specific and observed, not theoretical: pointing retrieval at an unfiltered corpus lets stale entries poison results. One team's first attempt searched a whole knowledge base and hallucinated, because outdated articles were in scope; the fix was converting only handpicked, actively maintained articles. Curation here is not fastidiousness — it is what makes the corpus usable at all. A related demonstration is worth keeping: an agent asked to check its own memory files for duplication reported none, and then, once steered to compare intent rather than text, found it had written the same memory repeatedly and was ignoring it.

Also: the cq exchange from Mozilla AI, with its MCP server, skill plugin and local SQLite store; Obsidian; G-Brain; ZeroEntropy.

Assay's insight-routing rule already names three levels — per-project memory, a central tracker for systemic and process insights, and the public canonical methodology spec — and the boundary between private and public is enforced hard. What is missing is the nomination half. There is no defined path by which a lesson that has proved out across several sessions gets promoted into the shipped skill bundle or the spec. Promotion happens today because a human notices, which is exactly the discretionary step the tiering is meant to replace.

The standing rule here is refresh, don't remember: re-fetch rather than trust recall, and bind every decision to this cycle's reads. The typed-unit schema is a direct upgrade path — evidence links plus confirm-and-flag semantics turn that discipline into recorded state, so a stale entry becomes detectable rather than merely distrusted by policy. The raw ingredient is already there in the findings register, where each entry is a discrete file with an incident behind it; what it lacks is the confidence field and the confirm-on-use write.

An agent finding no duplicates by text and many by intent describes Assay's memory files exactly: an index of prose lines accumulated over months, with no deduplication pass running against them at all. And the stale-knowledge-base result is the operational warning — any future move to point roles at an unfiltered corpus, whether raw streams, raw transcripts or the whole findings directory, should expect stale-entry poisoning and should promote a curated subset instead.

• Whether anyone implements nomination-and-graduation as mechanism rather than as a described policy — the signal that this is a pattern rather than one team's discipline.
• Whether confidence scores accumulated from agent confirm and flag signals turn out to be well calibrated, or drift into everything-scores-high.
• No Agentic AI Summit 2026 evidence appeared for this concept in the initial scan; a second event corroborating it would move the status toward established.