sqlite-muninn 0.3.1

HNSW vector search + graph traversal + GGUF embeddings + Node2Vec for SQLite

sqlite-muninn

Odin's mythic raven of Memory.

Huginn and Muninn fly each day over the wide world.
I fear for Huginn that he may not return,
yet I worry more for Muninn.

- Poetic Edda (Grimnismal, stanza 20)

Odin fears losing Memory more than Thought.

This project aims to build agentic memory and knowledge graph primitives for sqlite as a native C extension. It is an advanced collection of knowledge graph primitives like Vector Similarity Search, HNSW Indexes, Graph database, Community Detection, Node2Vec capabilities and loading GGUF models via llama.cpp integration.

Documentation | GitHub

Package Index	Published Version	Downloads
PyPI
npm

Features

• HNSW Vector Index - O(log N) approximate nearest neighbor search with incremental insert/delete
• Graph Traversal - BFS, DFS, shortest path, connected components, PageRank on any edge table, dbt syntax graph node selection.
• llama.cpp native models; Load and use GGUF LLM models natively in sqlite.
• Centrality Measures - Degree, betweenness (Brandes), and closeness centrality with weighted/temporal support
• Community Detection - Leiden algorithm for discovering graph communities with modularity scoring
• Node2Vec - Learn structural node embeddings from graph topology, store in HNSW for similarity search
• Zero dependencies - compiles to a single .dylib/.so/.dll
• SIMD accelerated - ARM NEON and x86 SSE distance functions

Research References

Feature	Paper
HNSW	Malkov & Yashunin, TPAMI 2020
MN-RU insert repair	arXiv:2407.07871, 2024
Patience early termination	SISAP 2025
Betweenness centrality	Brandes, J. Math. Sociol. 2001
Leiden community detection	Traag, Waltman & van Eck, Sci. Rep. 2019
Node2Vec	Grover & Leskovec, KDD 2016
SGNS	Mikolov et al., 2013

License

MIT. See LICENSE.