Title: MemGraphRAG: A Memory-Driven Multi-Agent Architecture for Graph Retrieval-Augmented Generation

Abstract

Retrieval-Augmented Generation (RAG) has emerged as a critical technique for reducing hallucinations in Large Language Models (LLMs) by integrating external knowledge sources. While traditional RAG performs well on straightforward queries, it faces significant challenges when processing large-scale, unstructured datasets characterized by highly fragmented information. To address this, Graph-based RAG (GraphRAG) utilizes knowledge graphs to map structural relationships, thereby facilitating more thorough retrieval and supporting complex reasoning tasks.

Despite these advantages, current GraphRAG approaches often suffer from a lack of global context, as they depend on isolated, fragment-level extraction during graph construction. This limitation frequently results in graphs that are structurally disjointed, logically contradictory, or thematically inconsistent, ultimately undermining retrieval accuracy. To overcome these issues, we introduce MemGraphRAG, a novel framework that leverages a memory-based multi-agent system to guarantee the construction of high-quality graphs.

MemGraphRAG utilizes a collaborative network of agents backed by shared memory, which establishes a unified global context during the extraction phase. This design enables agents to dynamically resolve logical discrepancies and preserve structural connectivity across the entire corpus. Additionally, we introduce a memory-aware hierarchical retrieval algorithm specifically designed for the generated graphs. Comprehensive experiments across various benchmarks reveal that MemGraphRAG surpasses state-of-the-art baseline models while maintaining comparable efficiency. The source code is accessible at https://github.com/XMUDeepLIT/MemGraphRAG.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC