HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models

Summary

NeurIPS 2024 (cite Gutiérrez et al.; arXiv 2405.14831). Hippocampal indexing: LLM extracts a schemaless KG (neocortex), Personalized PageRank spreads activation from query entities (hippocampus) for single-step multi-hop retrieval. Up to +20% vs then-SOTA RAG on multi-hop QA; single-step HippoRAG matches or beats iterative IRCoT at 10–30× cheaper and 6–13× faster; combining them helps further. Also handles integration scenarios existing RAG misses. Discord/X is a 2026 hype recap by @N01ennn quoting their “How to be a Memory Engineer” X article (Stanford/Microsoft/Anthropic/Nvidia lenses); the substantial artifact is this 2024 paper. MIT code OSU-NLP-Group/HippoRAG. OSU + Stanford.

Keywords

hipporag · rag · knowledge-graph · pagerank · neurips-2024 · osu · x

Topics

RAG, knowledge graphs, long-term memory

Research notes

• Primary: arxiv abs 2405.14831 (NeurIPS 2024). Discord/X https://x.com/N01ennn/status/2085758307495498103 via fxtwitter (third-party recap quoting X article https://x.com/N01ennn/status/2083971749079581120, id 2083965152056033281). Code https://github.com/OSU-NLP-Group/HippoRAG (MIT). Distinct from papers_local 541 (MeMo, which cites HippoRAG). Retrieval framework, not a new hosted corpus, so no datasets_local row.