HippoRAG 2

A knowledge-graph RAG method built on a neurobiological analogy, and the strongest, most token-efficient graph method in the neutral graphrag-bench benchmark. The paper is From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (Gutiérrez, Shu, Qi, Zhou, Yu Su; OSU-NLP Group, ICML 2025; arXiv:2502.14802). It extends the original HippoRAG (NeurIPS 2024; arXiv:2405.14831), so this page covers the lineage.

The neurobiological idea (HippoRAG v1)

HippoRAG borrows the hippocampal indexing theory of human long-term memory and maps it onto a RAG pipeline:

• Neocortex → the LLM (semantics, reasoning),
• Hippocampus → a knowledge-graph (the structured index of what’s been seen),
• Parahippocampal regions → the retrieval encoder (linking query terms to graph nodes).

The corpus is turned into a graph by LLM OpenIE (open information extraction of entities/relations), and retrieval runs Personalized PageRank (PPR) over that graph from the query’s entry nodes. PPR does in a single step what iterative RAG (e.g. IRCoT) needs many LLM calls for — it spreads activation across connected facts, capturing multi-hop associations at once. HippoRAG reports up to 20% gains on multi-hop QA over prior RAG while being 10–30× cheaper and 6–13× faster than iterative retrieval.

What HippoRAG 2 adds

v1 bought multi-hop and associative power but, like other graph-RAG, lost ground on plain factual recall (the graphrag-bench tradeoff in miniature). HippoRAG 2 reconciles that with deeper passage integration — keeping full passages in the graph alongside the OpenIE triples — and more effective online use of the LLM during retrieval. The result: it beats standard vector RAG across factual, sense-making, and associative memory at once, including a reported +7% on associative memory over a state-of-the-art embedding model. The framing — “from RAG to memory” — positions it as non-parametric continual learning: new documents extend memory without retraining, the same accumulation argument the temporal-knowledge-graph / agent-memory-knowledge-graphs thread makes for agent memory.

Why it stands out here

This is the one method where an advocate paper and a neutral benchmark agree. The authors’ own claims are author-run, but graphrag-bench (which proposes no method) independently ranks HippoRAG 2 as the top graph performer on complex reasoning (~54% vs ~43% for vanilla RAG) and the most token-efficient graph method (≈10³ tokens/query) — an order of magnitude under LightRAG and ~40× under Microsoft GraphRAG‘s global mode. So where graphrag-bench warns that graphs are a costly, task-conditional bet, HippoRAG 2 is the data point showing the efficiency gap is a design problem, not intrinsic to graph RAG. Distinct from GraphRAG (community summaries for global sensemaking) and leanrag (linking summary nodes): HippoRAG’s mechanism is PPR over an entity+passage graph for multi-hop retrieval. Authors / OSU-NLP noted inline (thin).

Related

graphrag-bench · graphrag · from-local-to-global-graphrag · leanrag · lightrag · raptor · knowledge-graph · retrieval-augmented-generation · temporal-knowledge-graph · agent-memory-knowledge-graphs