Your RAG pipeline needs a hippocampus

What it does

HippoRAG 2 indexes external documents into a knowledge graph and retrieves context by ranking concept connections with Personalized PageRank, aiming to mimic human long-term memory rather than simple vector lookup. It handles both straightforward factual retrieval and multi-hop reasoning across complex narratives, and it incrementally updates its graph as new documents arrive. The system can run against OpenAI APIs or locally hosted vLLM instances.

The interesting bit

The framework deliberately separates cheap, fast online retrieval from heavier offline graph construction, and the authors claim it uses fewer resources than comparable graph-based systems like GraphRAG or LightRAG while still improving multi-hop retrieval. The neurobiological framing isn’t mere branding—the architecture is built around continual learning and associative memory rather than one-shot search.

Key highlights

• Persistent knowledge graph built via OpenIE, enabling multi-hop associative retrieval across disconnected passages
• Personalized PageRank over the graph finds context chains instead of relying on pure vector similarity
• Supports both OpenAI and local vLLM backends, plus an offline batch indexing mode the authors say is more than 3× faster than online serving
• Evaluated on continual-learning benchmarks spanning factual memory, sense-making, and associativity
• Built-in support for incremental updates and document deletion, tested as core operations

Caveats

• Embedding support is currently restricted to NV-Embed, GritLM, and Contriever
• The README repeatedly warns about GPU OOM errors during local deployment and advises manually clearing cached OpenIE results and knowledge graphs between experiment reruns
• Claims of lower resource usage versus GraphRAG and LightRAG are stated without specific numbers or tables in the README

Verdict

Try it if you need RAG that accumulates and connects knowledge over time rather than answering isolated questions. Look elsewhere if you want a plug-and-play vector search replacement with broad embedding model support.