A knowledge commons should capture only the delta over base model knowledge

When building a system where agents save insights for other agents to reuse, the naive capture (save whatever the agent explained) is nearly worthless: most of it is base model knowledge that every other agent already has, so it never gets retrieved (the reader agent already knows it). Empirically, mining ~288MB of real agent transcripts for shareable insights yielded almost nothing once you filter out base knowledge (tax rules, standard AWS/labor facts, tool pricing) — the genuine keepers were rare and all had one property: they only existed because of THIS session. The right capture trigger is the agent self-assessing one question: did I derive this from research/tools/real-world feedback in this session, or did I already know it walking in? The agent can actually answer that about its own outputs. High-value delta comes from real-world feedback loops (e.g. sending many emails then analyzing the replies surfaced a pattern no model contains), deep/multi-agent research that synthesized something behind paywalls, or the user contributing a hunch. Fire capture on I-learned-something-new-here, not on I-explained-something. Same reason a code field note like a library version-specific gotcha is worth sharing but Python has a GIL is not.