HippoRAG

HippoRAG is a RAG framework inspired by human long-term memory. It uses LLMs, knowledge graphs, and Personalized PageRank to help a system retrieve information associatively across documents. In this vault, HippoRAG matters as a memory/retrieval-first system that helps explain the bridge from rag toward agent-memory-systems.

Plain-language explanation

Classic RAG can retrieve passages that look similar to a question. But if the answer depends on connecting facts across several documents, simple chunk retrieval can fail.

HippoRAG tries to imitate a piece of human memory: the hippocampus. The idea is that memory is not just a box of documents; it is an index of associations. When you remember one thing, nearby related things become easier to recall.

Core mechanism

HippoRAG works in two broad phases:

1. Offline indexing — an LLM extracts a schemaless knowledge graph from passages. This graph acts like an artificial hippocampal index.
2. Online retrieval — the system extracts named entities from a query, links them to graph nodes, and runs Personalized PageRank from those nodes to find relevant neighborhoods.

Personalized PageRank is important because it lets the system spread attention through a graph from the query’s starting points. That means HippoRAG can retrieve across multiple hops in one retrieval process, rather than needing repeated query-retrieve-generate loops.

HippoRAG 2

HippoRAG 2 extends the original line from RAG toward memory. Its paper, From RAG to Memory, argues that vector retrieval alone is not enough to mimic human long-term memory. It focuses on three capabilities:

• factual memory — retrieving direct facts reliably;
• sense-making — integrating large or complex contexts;
• associative memory — connecting related information across documents.

The key claim is that graph-augmented retrieval should not improve associativity while damaging factual recall. HippoRAG 2 tries to improve all three together.

HippoRAG 2 reports a 7% improvement in associative memory tasks over state-of-the-art embedding models while also improving factual and sense-making capabilities. The paper frames this as non-parametric continual learning for LLMs.

Relationship to LLM Wiki

HippoRAG and llm-wiki are both concerned with compounding knowledge, but they choose different artifacts and belong to different families. HippoRAG is graph/retrieval-first (moving toward memory-first via HippoRAG 2), while LLM Wiki is artifact-first:

• HippoRAG’s central artifact is an associative retrieval graph using Personalized PageRank.
• LLM Wiki’s central artifact is a readable Markdown wiki.
• HippoRAG is optimized for retrieval and multi-hop association.
• LLM Wiki is optimized for legibility, editing, provenance, and long-term knowledge maintenance.
• HippoRAG’s update model is indexing new passages into a schemaless graph; LLM Wiki’s update model is extending existing pages with preserved prior knowledge.

For Synaptic Lattice, HippoRAG is conceptually important because it gives a technical language for “associative recall.” It helps explain how a future system might surface non-obvious links between notes.

HippoRAG (v1) was accepted at NeurIPS 2024. Its repository (OSU-NLP-Group/HippoRAG) has approximately 3.9k stars and an MIT license. ^[raw/papers/hipporag-neurobiologically-inspired-long-term-memory.md]

Practical interpretation for Synaptic Lattice

HippoRAG is not the first layer to build. It becomes relevant after the vault has enough high-quality pages and sources. A future experiment could use HippoRAG-like graph traversal to suggest connections between wiki pages, but the suggestions would need to remain visually distinct from manual links.

This matters because the user wants both manual and inferred connections. HippoRAG-style retrieval could power inferred suggestions, but it should not silently rewrite the human-facing wiki.

Related

• agent-memory-systems — broader family HippoRAG belongs to.
• rag — baseline retrieval approach HippoRAG extends.
• knowledge-graph — graph representation used for associative retrieval.
• lightrag — related graph/retrieval-first system.
• lightrag-vs-hipporag — comparison between both approaches.
• memory-retrieval-patterns — pipeline techniques (BM25, RRF, re-rank, MMR) used alongside graph traversal.
• memory-retrieval-patterns-vs-wiki-query — page-level PageRank vs Personalized PageRank distinction.

Sources

^[raw/papers/hipporag-neurobiologically-inspired-long-term-memory.md] ^[raw/papers/hipporag-2-from-rag-to-memory.md] ^[raw/repositories/osu-nlp-group-hipporag-github.md]