Research map / Large language models / Retrieval and knowledge
HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HippoRAG, a neurobiologically inspired RAG framework that builds an LLM-extracted open knowledge graph as a hippocampal index and uses Personalized PageRank for single-step multi-hop retrieval.
Problem
Current RAG methods encode passages in isolation and struggle with knowledge integration across passages, requiring costly iterative retrieval for multi-hop QA and failing on path-finding questions.
Approach
Mimic hippocampal indexing: use an instruction-tuned LLM to extract an open schemaless KG from passages (neocortex/hippocampal index), link query entities via retrieval encoders, and run Personalized PageRank from query nodes to rank passages in a single retrieval step, with node specificity and synonymy edges.
Key results
- Single-step HippoRAG (ColBERTv2) improves R@5 by about 3 points on MuSiQue and about 20 points on 2WikiMultiHopQA over ColBERTv2 (average R@5 72.9 vs 65.6).
- IRCoT + HippoRAG (ColBERTv2) reaches average R@5 of 78.2 and average QA F1 of 51.7, the best among compared methods.
- Single-step HippoRAG is on par with or better than IRCoT while being 10-30x cheaper and 6-13x faster online.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HippoRAG instead of the strongest listed single-step baseline | ColBERTv2 retrieval · MuSiQue | R@5: 50.1 -> 51.9 | Helped |
| Use HippoRAG instead of ColBERTv2 | single-step retrieval · 2WikiMultiHopQA | R@5: 68.2 -> 89.1 | Helped |
| Use HippoRAG instead of ColBERTv2 | single-step retrieval · HotpotQA | R@5: 79.3 -> 77.7 | Hurt |
| Use HippoRAG instead of ColBERTv2 | QA reader · MuSiQue | F1: 26.4 -> 29.8 | Hurt |
| Use HippoRAG instead of ColBERTv2 | QA reader · 2WikiMultiHopQA | F1: 43.3 -> 59.5 | Helped |
| Use HippoRAG instead of ColBERTv2 | QA reader · HotpotQA | F1: 57.7 -> 55.0 | Hurt |
| Add node specificity (ablation) | HippoRAG · MuSiQue | R@5: 50.2 -> 51.9 | Helped |
| Add node specificity (ablation) | HippoRAG · 2WikiMultiHopQA | R@5: 88.8 -> 89.1 | No clear effect |
| Add node specificity (ablation) | HippoRAG · HotpotQA | R@5: 73.7 -> 77.7 | Helped |
| Add synonymy edges (ablation) | HippoRAG · 2WikiMultiHopQA | R@5: 85.6 -> 89.1 | Helped |
| Add synonymy edges (ablation) | HippoRAG · MuSiQue | R@5: 50.2 -> 51.9 | Helped |
| Add synonymy edges (ablation) | HippoRAG · HotpotQA | R@5: 75.7 -> 77.7 | Helped |
Surprising findings
- Replacing GPT-3.5 with open Llama-3.1-70B-Instruct for OpenIE matches or slightly exceeds GPT-3.5 on QA and retrieval, while Llama-3.1-8B is competitive on most datasets.
- Adding neighborhood nodes of query nodes without PPR hurts performance relative to using query nodes alone (e.g., MuSiQue R@2 25.4 vs 37.1).
Limitations
- All components are used off-the-shelf without fine-tuning; NER and OpenIE errors account for most failures.
- Scalability to much larger corpora has not been empirically validated.
- Entity-centric design ignores contextual cues, causing concept-context tradeoff errors (about 48% of MuSiQue errors are NER limitations).
Experimental setup
Datasets: MuSiQue 2WikiMultiHopQA HotpotQA
Benchmarks: MuSiQue 2WikiMultiHopQA HotpotQA
Metrics: recall@2 recall@5 all-recall@2 all-recall@5 exact match F1
Baselines: BM25 Contriever GTR ColBERTv2 RAPTOR Proposition (Propositionizer) IRCoT
Builds on: IRCoT Personalized PageRank ColBERTv2 Contriever
Closest papers by meaning
- From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025 · similarity 0.92)
- NeuroPath: Neurobiology-Inspired Path Tracking and Reflection for Semantically Coherent Retrieval (NeurIPS 2025 · similarity 0.88)
- Human-inspired Episodic Memory for Infinite Context LLMs (ICLR 2025 · similarity 0.87)
- Adapting to Non-Stationary Environments: Multi-Armed Bandit Enhanced Retrieval-Augmented Generation on Knowledge Graphs (AAAI 2025 · similarity 0.85)
- PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.85)
- PaceLLM: Brain-Inspired Large Language Models for Long-Context Understanding (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.