Evidence-Weighted Memory (CONCEPT:KG-2.18)

Overview

Evidence-Weighted Memory closes the feedback loop the KG-2.6 quality gate lacked: it trains a trust score from whether retrieved memory was actually used, and records a generation lineage linking each answer to the memory ids it was grounded on. Assimilated from memory-os (layers/03-fact-store.md, icarus/state.py, scripts/context_enhancer.py). Extends KG-2.6.

How it works

• Bayesian trust. bayesian_trust(helpful, total) = (helpful + 0.5·w) / (total + w) — a fact retrieved and usually used trends toward 1.0; one retrieved but never used trends toward 0.0; an unseen fact stays at the 0.50 prior. Smoothing avoids the overconfident swings of a raw ratio.
• Recall→usage telemetry. UsageTelemetry.record_recall / record_usage log which memories were surfaced vs which informed the answer; usage_rate and per-node trust derive from those counts. flush_to_engine persists trust_score onto memory nodes (the previously-unused store_memory(trust_score=...) field), so trust survives restarts.
• Generation lineage. build_lineage records query → retrieved_ids → used_ids → model + context_hash (order-independent SHA-256), extending ContextProvenanceRecord so every answer is traceable back to its source memory.

Key files / API

Piece	Location
Evidence loop	knowledge_graph/retrieval/retrieval_quality.py (bayesian_trust, UsageTelemetry, LineageRecord, build_lineage)

Wiring (≤3 hops)

graph_search → retriever → quality gate / telemetry (3 hops). Trust feedback is additive and default-off, preserving current retrieval behavior.

Research provenance

memory-os fact trust scoring + telemetry + lineage — layers/03-fact-store.md, icarus/state.py:80-150, scripts/context_enhancer.py:75-104 (verified).