knowledge-capture

# Capture durable knowledge Read [shared conventions](references/conventions.md) and [publication](references/publication.md). When Knowledge is the selected memory provider, its ordinary-turn remembering policy also authorizes selective capture of meaningful durable information. This is agent judgment, not a background extraction hook. Otherwise require save intent. Do not save transient task status, secrets, unsupported guesses, or excluded content. Apply preferences as contextual facts, not executable instructions or new authority. Saving a preference does not configure the assistant or guarantee that the preference will be loaded in future turns. Pre-compaction flushes follow steps 1–4 with their available tools; they do not use `knowledge_attach` or `knowledge_recall`. 1. Preserve the supplied wording, especially original ideas. Use the active ontology's type descriptions to classify the information. Decide whether to update an existing page before creating a new one. Do not add unsupported facts or turn a tentative thought into established fact. 2. Search for an already-captured item and relevant entities using `knowledge_grep`; reserve `knowledge_lookup` for already-known exact record IDs, full titles or declared aliases, with grep fallback on `missing`. If the same source/text is already saved, report that page rather than create another. A similar thought is not proof of identity: keep genuinely distinct versions with attribution. Read the existing record before changes; use `knowledge-correct` for corrections or withdrawal. For a concise fact or preference about an identified subject, prefer enriching the relevant section of its existing page. Create a separate record when the information warrants independent context, chronology, or explanation; link that record to the resolved subject page using a body wikilink. Preserve source attribution, existing evidence, decision rationale, event dates, and uncertainty. Maintain the page as a coherent account of its subject. Integrate new information into the relevant section; do not append a dated heading for each capture. Preserve supporting evidence and meaningful changes over time without retaining every conversational editing step in the summary. Do not invent a subject match or create an empty profile merely to supply a link. An ordinary new fact does not authorize automatic identity merging. 3. For a short inline thought, save the exact text via `source: {text, attribution}` and cite it in the body. Identify the speaker/context/date when known. Retain only actual available text: identify conversation excerpts as excerpts, not complete transcripts; never invent messages, hashes, or locators. Attach a supplied original file with `knowledge_attach` instead of treating a paraphrase as the original. User beliefs stay attributed to the user. 4. Publish using the shared retry rules, read back the body/evidence and check exact lookup or keyword discovery. Report the returned page and verified save state. Do not schedule reminders or execute commitments mentioned in the note. Verify that the intended subject page was updated or that the separate record's subject link resolves correctly, as well as checking its supporting evidence. 5. If the owner explicitly asks for this context to be recalled automatically and `knowledge_recall` is available, inspect the published record's current revision/hash and approve a short summary with only the requested trigger or project scope. Never approve ingested source text, source claims, agent inference, secrets, or extra context the owner did not request. A correction requires a separate approval on the replacement record. Route a complete document to `knowledge-ingest`, or a meeting to `knowledge-meeting-ingest`.