Skip to content

LlmWikis knowledge page

MATM Retrieval

MATM retrieval should return scoped, reviewed, source-backed agent experience only when it is safe and useful. Similarity alone is not enough; retrieval must filter by scope, status, source, sensitivity, authorization, conflict state, and deprecation state.

Retrieval Steps

  1. Plan the retrieval: task, consumer identity, project/tenant, allowed scopes, allowed kinds, and required evidence.
  2. Filter hard: authorization, scope, sensitivity, review status, current/deprecated state, and source availability.
  3. Retrieve candidates: lexical, structured, graph, vector, or trajectory-state search can be combined.
  4. Rank responsibly: relevance, task state, producer reliability, freshness, evidence, conflict status, and consumer fit.
  5. Abstain when needed: missing authorization, weak evidence, only hypotheses, stale/deprecated records, unresolved conflicts, or cross-scope leakage risk.
  6. Cite results: return record ids, source refs, curation report ids, warnings, and deprecation/conflict notices.

Trust-Aware Ranking Signals

Signal Positive use Warning
Scope match Project record used inside same project. Cross-project records require explicit approval.
Evidence quality Reviewed source refs and artifact hashes. Evidence record is not automatically the claim.
Producer reliability Past useful trajectories can influence rank. Do not turn reliability into unconditional authority.
Freshness Recent current records rank over stale ones. Fresh wrong memory is still wrong.
Conflict/deprecation state Warnings and redirect to replacement records. Do not hide credible conflicts.

Non-Normative API Example

The starter bundle includes llm-wiki/matm/examples/reference-api.openapi.yaml as an external-runtime example. It is not a live LLMWikis API.

POST /matm/retrieval-requests
request:
  consumer_agent: "consumer.demo"
  task: "repair checkout retry"
  allowed_scopes: ["project", "agent_team"]
  required_metadata: ["source_refs", "status", "scope", "authorization_context"]
response:
  abstained: false
  citations:
    - record_id: "mem_demo_team_procedure_001"
      warning: "project-specific validation required before use"