MemGraphRAG
A memory-enhanced GraphRAG framework accepted at a 2026 data-mining conference that organizes retrieval into three connected layers, an abstract schema, extracted facts, and source passages, with conflict detection between facts and their supporting evidence.
MemGraphRAG structures retrieval-augmented generation around three linked memory layers: a schema layer of abstract ontology triples, a fact layer of concrete relations extracted from a corpus, and a passage layer of the original source text, connected by bidirectional links so retrieval can move between abstract patterns and their concrete evidence. Ontology induction abstracts recurring facts into reusable schemas and filters out low-frequency patterns, and a conflict-aware construction step detects hard conflicts between facts and passage evidence and resolves connected groups of conflicts before building the final memory-derived graph.
Retrieval combines embedding similarity with personalized PageRank over the constructed graph, and a batch question-answering mode records answers, retrieved documents, scores, latency, and token usage for evaluation. It requires Python 3.10 or later, and its online pipeline uses an OpenAI-compatible LLM endpoint together with a local Hugging Face embedding model.
The project is maintained by the same research group behind several related graph-retrieval papers, including a GraphRAG benchmark and survey.