Context. kg_promotion.PromoteTriplesToFacts is a pure per-row read-Cypher→Fact pass-through with no global corpus view — exactly the "isolated local extraction" failure mode MemGraphRAG diagnoses (GraphRAG raises recall 84.3 vs 71.8 but drops relevance 38.5 vs 62.9; filtering ~40% low-frequency triples improves accuracy). Adding a candidate→stable schema-promotion gate keyed on extraction frequency is the lowest-architecture-disruption attack on thematic irrelevance.
Proposal. Add a Unified Schema Filtering τ-gate before the per-row fact_store.pin().
- Aggregate
result.rows by schema signature (predicate; + subject/object kind once type nodes exist) into a frequency map.
- Add
tau: int = 1; only rows whose schema frequency ≥ τ are pinned; the rest dropped or pinned with a 'candidate' lineage flag.
- Emit the candidate/stable decision into the existing lineage dict. Keeps the read-only-Cypher contract intact — pure additive gate.
Where.
src/stargraph/stores/kg_promotion.py:139-198 — per-row pin loop; gate before fact_store.pin() :197; interim in-process Counter over result.rows :139; lineage dict :186-193.- Durable frequency source = Type-node frequency column from the tri-view schema feature (linked issue).
Acceptance criteria.
Caveat. The interim Counter only sees rows the filter query returns, not the true global corpus — a weaker approximation of the paper's global memory until frequency is materialized on Type nodes.
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.
kg_promotion.PromoteTriplesToFactsis a pure per-row read-Cypher→Fact pass-through with no global corpus view — exactly the "isolated local extraction" failure mode MemGraphRAG diagnoses (GraphRAG raises recall 84.3 vs 71.8 but drops relevance 38.5 vs 62.9; filtering ~40% low-frequency triples improves accuracy). Adding a candidate→stable schema-promotion gate keyed on extraction frequency is the lowest-architecture-disruption attack on thematic irrelevance.Proposal. Add a Unified Schema Filtering τ-gate before the per-row
fact_store.pin().result.rowsby schema signature (predicate; + subject/object kind once type nodes exist) into a frequency map.tau: int = 1; only rows whose schema frequency ≥ τ are pinned; the rest dropped or pinned with a'candidate'lineage flag.Where.
src/stargraph/stores/kg_promotion.py:139-198— per-row pin loop; gate beforefact_store.pin():197; interim in-processCounteroverresult.rows:139; lineage dict :186-193.Acceptance criteria.
Caveat. The interim
Counteronly sees rows the filter query returns, not the true global corpus — a weaker approximation of the paper's global memory until frequency is materialized on Type nodes.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.