HippoRAG

HippoRAG is an open-source retrieval and long-term memory framework developed by the Ohio State University NLP Group. Inspired by the hippocampal indexing theory of human memory, HippoRAG enables Large Language Models (LLMs) to perform complex, multi-hop associative recall across knowledge corpora in a single retrieval step using knowledge graphs and Personalized PageRank (PPR).

graph TD
    Corpus[Unstructured Document Corpus / Memories] --> OpenIE[Open Information Extraction via LLM]
    OpenIE --> KG[Hippocampal Knowledge Graph<br/>Entity Nodes & Relationship Triples]
    Query[Multi-Hop User Query] --> QueryNER[Query Named Entity Recognition]
    QueryNER --> SeedNodes[Graph Seed Nodes]
    SeedNodes --> PPR[Personalized PageRank Algorithm]
    PPR --> TopPassages[Ranked Associative Passages]
    TopPassages --> LLM[LLM Response Generation]

Neurobiological Inspiration

Traditional RAG systems mimic the neocortex by relying on dense vector embeddings for semantic similarity, which struggle with multi-hop associative reasoning without slow, iterative multi-turn retrieval loops.

HippoRAG models human brain architecture:

• Neocortex (LLMs & Embeddings): Extracts structured triples and forms representations.
• Hippocampus (Personalized PageRank & Graph Index): Acts as an indexing engine connecting distinct entities across disparate documents, allowing pattern separation and associative pattern completion in a single algorithmic pass.

Key Characteristics

• Single-Step Multi-Hop Retrieval: Discovers associative connections spanning multiple documents without iterative retrieval latency.
• High Efficiency: Significantly cheaper and faster at indexing and querying compared to iterative GraphRAG or iterative agentic search workflows.
• Continual Learning: Accommodates incremental updates to the knowledge graph without requiring retraining or full corpus re-embedding.
• Open Source & Local: Published under the MIT license, available via PyPI (pip install hipporag), and fully runnable on local hardware using local LLMs (e.g., via vLLM or Ollama).

Python Quickstart

from hipporag import HippoRAG

# 1. Initialize HippoRAG instance
hipporag = HippoRAG(
    llm_model="meta-llama/Llama-3.3-70B-Instruct",
    embedding_model="nvidia/NV-Embed-v2"
)

# 2. Index corpus of documents or agent memories
documents = [
    "Alice founded Quantum Dynamics in 2022 in Geneva.",
    "Quantum Dynamics specializes in topological quantum computing chips.",
    "The Swiss National Science Foundation awarded Geneva-based quantum startups 10M CHF."
]
hipporag.index(documents)

# 3. Perform associative multi-hop retrieval
query = "What funding did Alice's research domain receive in Switzerland?"
results = hipporag.retrieve(query, top_k=2)
for doc in results:
    print("Retrieved context:", doc)

Related Concepts

• LLM and Agent Memory - Foundational agent memory concepts and cognitive taxonomy.
• Cognee - Topological knowledge graph memory engine.
• Graphiti and Zep - Temporal knowledge graph engine for dynamic agent memory.