HippoRAG is a retrieval-augmented generation framework inspired by how the human hippocampus indexes memory. Instead of matching isolated text chunks, it builds a knowledge graph over your corpus and uses Personalized PageRank to retrieve across documents — enabling the multi-hop reasoning that standard vector RAG fumbles. The project has about 4,100 stars on GitHub and is actively developed in 2026.
Published at NeurIPS 2024 by Ohio State’s NLP group, HippoRAG extracts entities and relations with an LLM during indexing, links them with synonymy edges from embeddings, then seeds PageRank with the entities mentioned in a query so retrieval spreads through the graph the way associative memory does. The payoff shows on multi-hop question answering (MuSiQue, 2WikiMultiHopQA, HotpotQA), where it outperforms strong RAG baselines while being cheaper and faster online than iterative retrieval agents. The framework supports OpenAI and local models, multiple embedding backends, and has grown to include HippoRAG 2 improvements. For applications like enterprise knowledge bases, scientific literature and legal research — where answers require integrating facts scattered across many documents — it offers a principled, well-evaluated upgrade path from naive chunk retrieval.
The framework’s indexing cost is front-loaded — building the graph requires LLM extraction over the corpus — but online queries then run a single fast retrieval step, an attractive trade for knowledge bases that are queried far more often than they change. It integrates with standard vector stores for the embedding side, exposes its graph for inspection, and the repository includes the exact evaluation harness from the paper for MuSiQue, 2WikiMultiHopQA and HotpotQA. HippoRAG 2 further improves passage integration and recognition memory within the same codebase.
HippoRAG is free and open source under the MIT license; you supply the LLM/embedding endpoints it calls.
HippoRAG is for RAG builders and researchers whose questions span many documents — knowledge bases, literature review, support corpora — and who need retrieval that connects facts, not just matches text.