Semantic Memory
Version: 1.0.0 | Last updated: 2026-07-16
Purpose
Serve governed factual and conceptual knowledge with source-level attribution and temporal validity.
Why
Semantic memory becomes dangerous when extracted claims lose their source, scope, qualifiers, or effective dates.
How
Ingest authoritative sources through validation and provenance capture. Represent claims with subject, predicate, object/value, qualifiers, valid time, source version, confidence, ACL, and supersession. Keep extracted claims linked to source spans. Resolve conflicts by documented source authority and time; return multiple claims or abstain when unresolved. Rebuild derivatives after corrections.
Tradeoffs
Normalized claims support precise retrieval and conflict handling but require extraction and stewardship. Source chunks are simpler but make entity and temporal reasoning harder.
Anti-patterns
- Treating embedding similarity as factual confidence.
- Flattening conflicting claims into one “canonical” answer without policy.
- Dropping qualifiers during extraction.
Enterprise Considerations
Assign domain owners and authority tiers. Version taxonomies, preserve lineage, and audit corrections affecting regulated decisions.
Checklist
- Claims retain source spans, qualifiers, and effective time.
- Authority and conflict policies are explicit.
- Corrections propagate to indexes and answers.
- Unsupported claims trigger abstention.
References
Changelog
- 1.0.0 — 2026-07-16: Initial production standard.