Wu, Xiang, Tang, Chen, Zhang and Su (2026) — ARXIV-2606.00610, read as NOTE-249. KDD 2026.
Key takeaways
- The pilot study measures a trade-off in other people’s systems, which is why it is the part worth keeping. On G-Medical, GraphRAG pipelines beat vanilla RAG on recall (GFM-RAG 84.3% vs 71.8%) and lose badly on relevance (38.5% vs 62.9%), and the noisier context costs generation accuracy. The authors are proposing a GraphRAG system, so this is a finding against their own family’s interest.
- Forty per cent of a constructed graph is free to delete. Filtering triples by schema frequency and dropping the bottom 40% moves accuracy from 64.85% to 65.28%. The paper reads that as evidence of “thematically irrelevant noise”; the number itself supports a cost claim more strongly than an accuracy one — 0.43 points is small, and nothing lost across a 40% reduction is the result.
- Three named failure modes of isolated local extraction: thematic irrelevance (no global view of corpus theme), logical inconsistency (mutually exclusive, temporal and granularity conflicts between independently extracted chunks) and structural fragmentation (entities duplicated across disconnected subgraphs because coreference is never resolved globally).
- The system: a three-tier global memory (ontology, fact, passage) with
three agents (extract, detect conflicts, resolve conflicts), schema
promotion gated on frequency
Freq(s) ≥ τ, and Personalized PageRank at retrieval. Headline average 59.68 against LazyGraphRAG’s 44.97, and retrieval latency 0.061 s against LightRAG’s 11.052 s.
Standing in the anthology
Filed for SOTA-308, which rests on the pilot study. No practice rests on MemGraphRAG itself, and the reading says why at length: single runs, no seeds, the authors’ own system against baselines they ran, and a transferability claim whose deltas are sub-point.
That last one is worth naming here because it is the paper’s weakest passage and its strongest verbs. §5.4 replaces four systems’ graph constructors with this one and reports HippoRAG 51.07 → 51.78 and MS-GraphRAG 43.75 → 44.21 — under a point, one run each — as “consistent improvement” that “substantially strengthens the effectiveness of existing retrievers”. Having just filed SOTA-307 on seed variance in a neighbouring field, the record is not in a position to take sub-point single-run deltas at their framing.
The corpus holds SOTA-269 — keep a knowledge base outside the weights — and nothing about how to build or evaluate one. This is the first document in that space, and it arrives as a caution rather than a method.