knowledge-recall skill

Recall saved knowledge

Read shared conventions. This workflow is read-only.

and pages whose exact stored identity is not already known: literal words, all by default; try a concise variant, any, or phrase when useful. No embeddings, regex, graph traversal or server-side synthesis is available. Narrow by ontology type/kind when appropriate; use knowledge_records and pagination for an inventory, not a ranked search pretending to be exhaustive. A hit with matches: ["superseded"] is the current replacement of an earlier record the query matched; via names that earlier revision. The words matched via, not the returned record.

1. Search with knowledge_grep({query}), including for named people, projects

full title or declared alias. It does not perform keyword search. If it returns missing, fall back to knowledge_grep before concluding no relevant record exists. Preserve ambiguous as ambiguity and inspect candidates rather than choosing one.

1. Use knowledge_lookup({query}) only for an already-known exact record ID,

the answer. Identity, aliases, domain fields and linked pages are included; request includeEvidence: true for source locators, excerpts and attribution support. Connection rows carry each claim's text, day, linked pages and source. Follow body/evidence/connection pagination as needed. Compare the hit's revision and 16-character hash with the readback's revision and hash prefix. Restart on changes rather than combine incompatible revisions.

1. Inspect relevant hits with knowledge_record. Read enough context to support

linked to the dedicated Knowledge record page using the shared conventions' chat-link contract. Retain relevant source locators and propagate original evidence, not only the generated summary. Label synthesis and disagreements. Use lifecycle history filters explicitly for historical questions; do not present superseded material as current.

1. Answer from the retrieved passages. In Control UI chat, cite readable titles

Follow result offsets with unchanged filters and, for keyword search, compare index.snapshot. If an API pagination bound prevents exhausting results, narrow the query or state the coverage limit. Do not claim the vault lacks something solely because a short ranked list did not contain it.

1. Distinguish no matches from inaccessible, invalid, incomplete or failed reads.

Recall itself is read-only. If the selected Knowledge memory policy calls for remembering genuinely new durable information, use knowledge-capture separately. Do not re-save retrieved summaries as new evidence or enrich entities merely because they appeared in an answer.

Temporal questions

Use knowledgechronicleday for a day/week, knowledgechroniclesince for an occurrence range, knowledgetimeline for an entity's dated records or a chosen timeField (group: "source" gives the entity's Timeline: one row per dated source with its note, filtered by the source's date; claims dated apart from their source keep their own rows), and knowledgechroniclelastseen for latest dated context. Resolve an entity to a stable page ID first. Dated mentions do not prove attendance. For keyword/inventory queries add since/until and timeField to knowledgegrep or knowledgerecords. occurredat answers when something happened; recordedat and updatedat answer when records were created/changed. Never substitute one for the other. Resolve relative query dates with the user's timezone, preserving explicit offsets; day queries accept an IANA timezone. Source date-only values stay date-only. Follow nextOffset with unchanged filters and returned snapshot. Temporal eligibility="includeinactive" includes superseded/expired material, never withdrawn records or a reconstruction of past system state. Last-seen uses occurredOnOrBefore for the occurrence cutoff; do not pass asOf or historical. Use includeSources=true for deduplicated source identities on temporal pages. Markdown-only timelines and undated records are not inferred, and no matches do not prove absence.

Focused source reading

Use knowledgeevidenceread with the record id and observed path/sha256 to read an exact retained UTF-8 passage. Supply inclusive startLine/endLine when known. Follow nextCursor with unchanged id/path/hash and omit or repeat line bounds. A response without nextCursor completes the passage. In contrast, record.nextEvidenceOffset only completes the citation list; excerptTruncated still requires source reading. The tool verifies bytes and line bounds, not the truth of an interpretation. Use knowledge_attachment for binary originals. Never infer complete review from an excerpt or a successful structural verifier.

Compare snapshot values only across pages with unchanged query filters and eligibility.