MemGraphRAG: Xiamen University's Memory-Based Multi-Agent System Rebuilds GraphRAG

Forum topic · 小凯 · 2026-06-15

Summary

MemGraphRAG, a KDD 2026 paper from Xiamen University and Jilin University, diagnoses a core flaw in existing GraphRAG systems: each document chunk is extracted independently with no global perspective, producing knowledge graphs plagued by off-topic noise, logical contradictions, and fragmented structure. The system addresses this with a three-layer global memory architecture (ontology, fact, and passage layers with bidirectional schema-instance and fact-evidence indexing) and a three-agent collaboration loop (extraction, conflict detection, and conflict resolution) that denoises topics and resolves contradictions during offline indexing. Online retrieval uses memory-guided hierarchical filtering, structure-aware node initialization with hub suppression, and personalized PageRank, achieving 0.061-second response times. Using only GPT-4o-mini, MemGraphRAG reaches 59.25% average accuracy across HotpotQA, 2Wiki, MuSiQue, and G-Bench, beating the LinearRAG SOTA baseline by 2.1 points, with 74.6% retrieval recall. Key limitations include text-only support, costly offline construction, and reliance on LLM reasoning for conflict resolution.

MemGraphRAG: Memory-Based Multi-Agent System for Graph RAG

> Paper: MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation > Venue: KDD 2026 | Institutions: Xiamen University × Jilin University > arXiv: 2606.00610 | Code: https://github.com/XMUDeepLIT/MemGraphRAG

TL;DR

MemGraphRAG diagnoses existing GraphRAG's "isolated amnesia": every document chunk is extracted independently with no communication, leading to thematically messy, logically contradictory, and structurally fragmented knowledge graphs. It combines a three-layer global memory (ontology–fact–passage) with three cooperating agents (extract → detect → resolve) to eliminate hallucinations during offline indexing, then uses memory-aware PPR retrieval for 0.061-second responses — exceeding SOTA accuracy by 2.1%.

The Problem: Three Core Defects of GraphRAG

Existing GraphRAG methods decompose global knowledge graph construction into countless independent local extraction tasks. Experiments confirm the noise: removing 40% of low-frequency triples barely changed accuracy.

| Defect | Manifestation | Consequence | |--------|---------------|-------------| | Topic irrelevance | Extracted triples drift from the central topic | Retrieval introduces noise | | Logical inconsistency | Contradictory facts in the same subgraph (Newton's birth year 1643 vs 1645) | Breaks semantic coherence | | Structural fragmentation | Key entities duplicated or scattered across disconnected subgraphs | No multi-hop traversal |

Core Techniques

1. Three-Layer Global Memory

Ontology Layer: schema patterns tracked with extraction frequency; patterns start "Pending" and are promoted to "Stable" only when corpus frequency exceeds threshold τ — a first-pass topic denoising filter.
Fact Layer: instantiated facts marked Active/Inactive; conflicts (e.g., Newton's birth year) are flagged, the resolved fact activated.
Passage Layer: original text passages for evidence tracing.

Two bidirectional index mechanisms bind the layers: Schema–Instance Alignment (every fact must follow a stable schema pattern) and Fact–Evidence Grounding (every fact links to its supporting passages).

2. Three-Agent Collaboration Loop

| Agent | Role | |-------|------| | Extraction agent (A_ext) | Processes chunks, extracting candidate schemas, facts, and source passages | | Conflict detection agent (A_det) | Monitors fact-layer updates; identifies redundancy, structural anomalies, and logical inconsistencies | | Conflict resolution agent (A_res) | Resolves conflicts using ontology constraints and passage-layer evidence |

Three conflict types, three strategies:

Mutually exclusive: Einstein born 1879 vs 1880 → keep the one with more evidence
Temporal: Biden vs Trump as US president → add time qualifiers (2021–2025 vs 2017–2021)
Granularity: birthplace LA vs USA → refine to "birth city" vs "birth country"

3. Memory-Guided Hierarchical Retrieval (0.061s)

Stage I: parallel multi-layer memory filtering with semantic similarity; falls back to standard RAG if no structured candidates.
Stage II: structure-aware node initialization — entity nodes use average fact similarity; type nodes apply hub suppression (relevance × 1/log(deg+1)) to prevent high-degree nodes like "Person" from over-propagating; passage nodes weight semantic relevance × damping × information density (high-IDF rare entities).
Stage III: personalized PageRank over the heterogeneous graph with λ=0.5; top-K passages and top-M entities feed the LLM.

Speed comes from doing all heavy computation (denoising, adjudication, structure unification) offline.

Experimental Results

Generation accuracy

| Method | HotpotQA | 2Wiki | MuSiQue | G-Medical | G-Novel | Avg | |--------|---------|-------|---------|-----------|---------|---------| | GPT-4o-mini zero-shot | 39.70 | 31.30 | 15.20 | 42.13 | 31.42 | 30.99 | | Standard RAG Top-5 | 60.30 | 45.40 | 32.00 | 61.07 | 48.35 | 47.97 | | HippoRAG2 | 67.20 | 57.90 | 38.30 | 64.85 | 56.48 | 55.79 | | GFM-RAG | 67.70 | 61.10 | 36.10 | 58.19 | 53.39 | 55.27 | | LinearRAG | 67.30 | 61.90 | 37.80 | 63.85 | 54.15 | 55.27 | | MemGraphRAG | 70.60 | 64.70 | 38.10 | 66.65 | 55.78 | 59.25 |

MemGraphRAG averages 59.25%, beating the strongest baseline LinearRAG by 2.10 points.

Retrieval quality and efficiency

Recall: 74.6% (exceeding RAPTOR by 10%+)
Response time: 0.061 seconds — comparable to standard RAG

Ablation study

| Variant | Accuracy | Loss | |---------|----------|------| | Full MemGraphRAG | 58.37 | — | | – unified schema filtering | 56.89 | -1.48 | | – global adjudication | 56.45 | -1.92 | | – hub suppression | 57.21 | -1.16 | | – information density term | 57.68 | -0.69 |

Unified schema filtering and global adjudication contribute the most — exactly the two core mechanisms of topic denoising and consistency maintenance.

Deep Dive: The Philosophy

1. Global perspective vs local extraction: graph quality depends not on individual extractor precision but on extractor collaboration — like Wikipedia as a product of many editors under shared norms. 2. Heavy offline, light online: quality is guaranteed offline; speed is realized online, mirroring search-engine architecture. 3. Architecture over compute: using only GPT-4o-mini, three specialized agents outperform one generalist — divide and conquer. 4. Dual meaning of "memory": global memory M (slowly built, globally consistent long-term memory) and memory-guided retrieval (fast, precise working memory).

Limitations

1. Text-only: no multimodal (image/audio) support yet. 2. Offline construction cost: high-quality memory requires heavy offline LLM calls; incremental updates for news/social media scenarios remain open. 3. Fixed frequency threshold τ: may filter legitimate schemas in sparse domains. 4. LLM dependency: conflict adjudication quality depends on the LLM's reasoning ability and knowledge.

Links

Paper: arXiv:2606.00610
Code: https://github.com/XMUDeepLIT/MemGraphRAG
Baselines compared: RAPTOR, HippoRAG2, LinearRAG, GFM-RAG
Benchmarks: HotpotQA, 2WikiMultiHopQA, MuSiQue, G-Bench (Medical/Novel)

> Key takeaway: the bottleneck of GraphRAG is not how large a graph can be built, but how consistent it can be. Only multi-agent collaboration over global memory — eliminating inconsistency at construction time — can turn a knowledge graph into a reliable external brain for LLMs.