AGENT SKILL

knowledge-capture

Save durable preferences, decisions, project facts, commitments, and lessons cheaply with knowledge_remember, keep the subject page's State current, or save short notes and page updates with attribution and readback. Use for explicit saves or agent-driven remembering when Knowledge is the selected memory provider.

License 许可
MIT
所属仓库
openclaw/openclaw
最近更新
2026-11-13
Stars
★ 39w
Forks
8.3w

01 前言 Front Matter

name
knowledge-capture
description
Save durable preferences, decisions, project facts, commitments, and lessons cheaply with knowledge_remember, keep the subject page's State current, or save short notes and page updates with attribution and readback. Use for explicit saves or agent-driven remembering when Knowledge is the selected memory provider.

02 资源文件 18

03 SKILL.md 正文

Capture durable knowledge

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 these steps with their available tools; they do not use knowledge_attach or knowledge_recall.

Save a claim

Most captures stop here and need no reference. Save a discrete claim (decision, preference, commitment, fact, event or belief) with knowledge_remember: one claim per item, short attribution, the supporting conversation excerpt and, only when you know it, the subject's exact page title or alias. An omitted subject is inferred when the text names exactly one entity page. It resolves subjects, reports exact repeats and never creates pages. Its published receipt is the save.

When the claim concerns an entity that has a page and changes what its State says (a decision, an outcome, a changed status, a resolved question), also update that page below in the same task. Claims alone leave its State stale. Continue below too to revise a page's prose or save a longer note.

Update a page or save a note

Read setup, subject pages and only the layout for the page's type (project, person, company, concept). Read publish before the write and writing before new prose. You don't need to reread them after compaction unless you are about to use them.

  1. Preserve the supplied wording, especially original ideas. Use the active ontology's type descriptions to classify the information. Do not add unsupported facts or turn a tentative thought into established fact.
  2. Search for an already-captured item and relevant entities with 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 or 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. Use knowledge-correct for corrections or withdrawal.
  3. Put the information on the most specific existing page it is about. Read the whole body first. Rewrite the State topic or type section it bears on in place so the page reads as current; never append a dated heading or a section per capture, and write no Timeline entries. Leave the rest of an already-current page alone, and keep every existing citation. Preserve decision rationale, event dates and uncertainty without retaining every conversational editing step. If the page has a written Timeline or dated section, retire it (retire dated sections).
  4. Create a separate record only when the information warrants independent context, chronology or explanation, and link it to the resolved subject page with a body wikilink. 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.
  5. Page bodies have a size budget (page budgets). If an edit returns state: "rejected", nothing was saved: follow that section's order, and save the fact with knowledge_remember meanwhile. Never truncate to fit.
  6. For a short inline thought, save the exact text via source: {text, attribution} and cite it in the body (sources). Identify the speaker, context and date when known. Identify conversation excerpts as excerpts, not complete transcripts; never invent messages, hashes or locators. Attach a supplied original file instead of treating a paraphrase as the original. User beliefs stay attributed to the user.
  7. Publish; on an uncertain or rejected write read recovery. Verify that the intended subject page was updated, or that the separate record's subject link resolves, and its evidence (verify). Report the returned page and verified save state. Do not schedule reminders or execute commitments mentioned in the note.
  8. If the owner explicitly asks for this context to be recalled automatically and knowledge_recall is available, inspect the published record's current revision and 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.

04 同仓库更多 Skills