Context. MemGraphRAG's Global Adjudication (consistency maintenance) needs conflict-driven resolution; its strongest ablation is that removing Conflict Resolution causes the biggest accuracy drop. In stargraph the Update/Delete delta primitives exist (UpdateDelta.replaces / DeleteDelta.replaces, memory.py:100-120) and apply_delta (sqlite_fact.py:192-222) is the only sanctioned write path into the semantic FactStore — but nothing produces conflict-driven deltas: consolidation only does intra-batch (subject,predicate) dedup and trusts rule output verbatim (embedding-similarity dedup deferred to Phase 3). Relates to #21 (promote memory stores out of POC, embedding dedup).
Proposal. Produce conflict-resolution deltas. On conflict (same subject/predicate, divergent object) emit UpdateDelta(replaces=[old_fact_id]); resolution policy = highest-confidence / most-recent. apply_delta needs no change (already UPDATE = unpin(replaces)+pin(new)).
Seat decision (needs-design). The MemoryStore Protocol's consolidate(rule) takes NO FactStore handle (memory.py:195; sqlite_memory.py:222-231 explicitly declines cross-store dedup for that reason). Realistic options:
- (a) a NEW free function alongside
kg_promotion taking (fact_store, fathom_adapter), emitting conflict deltas after scanning fact_store.query; or
- (b) broaden the
consolidate Protocol to accept an optional fact_store.
Where.
src/stargraph/stores/sqlite_memory.py:201-255 consolidate() + _classify_episode :258-326.src/stargraph/stores/sqlite_fact.py:192-222 apply_delta.src/stargraph/stores/_delta.py:42-62 _validate_delta_provenance.
Acceptance criteria.
Caveats. Temporal-conflict detection has no data: t_valid/t_invalid are declared in the Rel DDL (ryugraph.py:69) but NEVER written (add_triple ignores props — see the tri-view schema issue). Naive O(n²) scan is the scalability concern the paper flags.
Source: arXiv:2606.00610v1 — "MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation". Distilled from arXiv-research/2606.00610v1/analysis.md; file refs verified against current main by the analysis pass.
Context. MemGraphRAG's Global Adjudication (consistency maintenance) needs conflict-driven resolution; its strongest ablation is that removing Conflict Resolution causes the biggest accuracy drop. In stargraph the Update/Delete delta primitives exist (
UpdateDelta.replaces/DeleteDelta.replaces,memory.py:100-120) andapply_delta(sqlite_fact.py:192-222) is the only sanctioned write path into the semantic FactStore — but nothing produces conflict-driven deltas: consolidation only does intra-batch(subject,predicate)dedup and trusts rule output verbatim (embedding-similarity dedup deferred to Phase 3). Relates to #21 (promote memory stores out of POC, embedding dedup).Proposal. Produce conflict-resolution deltas. On conflict (same subject/predicate, divergent object) emit
UpdateDelta(replaces=[old_fact_id]); resolution policy = highest-confidence / most-recent.apply_deltaneeds no change (already UPDATE = unpin(replaces)+pin(new)).Seat decision (
needs-design). The MemoryStore Protocol'sconsolidate(rule)takes NO FactStore handle (memory.py:195;sqlite_memory.py:222-231explicitly declines cross-store dedup for that reason). Realistic options:kg_promotiontaking(fact_store, fathom_adapter), emitting conflict deltas after scanningfact_store.query; orconsolidateProtocol to accept an optionalfact_store.Where.
src/stargraph/stores/sqlite_memory.py:201-255consolidate()+_classify_episode:258-326.src/stargraph/stores/sqlite_fact.py:192-222apply_delta.src/stargraph/stores/_delta.py:42-62_validate_delta_provenance.Acceptance criteria.
UpdateDelta; resolution policy applied.source_episode_ids+rule_id(required by_validate_delta_provenance:54-60).Caveats. Temporal-conflict detection has no data:
t_valid/t_invalidare declared in the Rel DDL (ryugraph.py:69) but NEVER written (add_triple ignores props — see the tri-view schema issue). Naive O(n²) scan is the scalability concern the paper flags.Source: arXiv:2606.00610v1 — "MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation". Distilled from
arXiv-research/2606.00610v1/analysis.md; file refs verified against currentmainby the analysis pass.