MemGraphRAG: Multi-agent memory system beats traditional GraphRAG

GraphRAG enhances retrieval-augmented generation by incorporating knowledge graphs, but existing methods construct graphs from isolated fragments, often producing thematically inconsistent and logically conflicting structures. This degrades retrieval performance, especially for complex, large-scale corpora. To address this, researchers from multiple institutions propose MemGraphRAG, a novel framework that introduces a memory-based multi-agent system. The core innovation is a collaborative society of agents that share a unified global memory throughout the graph extraction process. This shared memory allows agents to maintain awareness of the entire corpus, dynamically resolve logical contradictions, and ensure structural connectivity across all extracted fragments. The result is a more coherent and context-aware knowledge graph.

Beyond graph construction, MemGraphRAG includes a memory-aware hierarchical retrieval algorithm tailored to leverage the improved graph structure. The algorithm uses the shared memory to guide traversal and aggregation, retrieving more relevant and complete information for complex queries. Extensive experiments on multiple benchmarks show that MemGraphRAG consistently outperforms state-of-the-art GraphRAG baselines while maintaining comparable efficiency. The work has been accepted to KDD 2026, and the code is publicly available on GitHub. This approach represents a significant step toward reliable, scalable knowledge integration for LLMs, reducing hallucinations and enabling deeper reasoning.