use-verifiable-memory

Use Verifiable Memory

Use the read-only Verifiable Memory Bank MCP tools. Do not present a search snippet or dossier summary as proof until its exact evidence packet has been checked.

Workflow

1. Call status first. Report an uninitialized, empty, degraded, or partially extracted bank before relying on it.
2. Start broad work with dossier. Use search to close gaps:
   • exact for names, identifiers, quotations, and distinctive phrases.
   • semantic for concepts expressed with different wording.
   • hybrid as the normal discovery mode.
3. For every material claim, call evidence with the result's segment_id. Use source with its source_id when provenance, document context, or source identity matters.
4. Check the packet before evaluating the claim:
   • Verify normalized_representation.sha256 and use the TextPositionSelector or TextQuoteSelector to reproduce the exact normalized span.
   • Treat offsets as coordinates in that named representation, not as byte, page, row, or media-time coordinates in the original object.
   • Use source_anchors when the adapter supplies page, row, line, cue, time, bounding-box, or byte coordinates.
   • Treat confirmation_eligibility as a trust and attribution gate only. When claim_evaluation is not_evaluated, compare the user's actual claim with the evidence text yourself.
5. Call chronology for sequences or changes over time. Keep material date, publication date, and acquisition date distinct. Preserve author and speaker attribution exactly, and do not assign unattributed text to a nearby person.
6. Synthesize only after verification. Organize the answer as Facts, Inferences, and Unknowns when uncertainty matters.

Evidence rules

• Treat a fact as supported only after directly comparing the concrete claim with the evidence text and confirming that the packet identifies a preserved source.
• Do not interpret confirms_claim: true as entailment. It is a compatibility field for source eligibility; claim_evaluation states whether the bank evaluated a concrete claim.
• Prefer canon evidence for factual conclusions. Label context and unresolved evidence explicitly.
• Cite the public object_ref or source ID plus segment ID. Include the representation identity and selectors when the reader may need to reproduce the exact span.
• State an inference separately and identify the facts it connects. Do not convert absence of evidence into evidence of absence.
• Put missing coverage, conflicting records, uncertain attribution, weak extraction, and unanswered questions under Unknowns.
• If evidence conflicts, show the conflict and relevant dates instead of selecting a convenient version silently.