SQLite Graph Memory

•
15
•

Provide associative graph-augmented retrieval and persistent memory over local Markdown notes using native wikilinks and SQLite.

SQLite Graph Memory is a lightweight Graph RAG memory layer and MCP server designed for coding agents such as Claude Code, Cursor, and Codex. Following the Andrej Karpathy LLM Wiki pattern, it turns any folder of Markdown notes or Obsidian vault into an associative knowledge base without requiring complex graph databases or heavy ETL pipelines. Dense vector search identifies initial entry points while human-curated wikilinks are traversed on the fly to surface deeply connected contextual notes that plain keyword or vector search miss, storing all recall telemetry and per-turn state in a simple, portable SQLite database.

主な機能

01Dynamic Graph RAG traversing handwritten wikilinks without separate extraction pipelines
02Single-file SQLite architecture requiring zero operational overhead or dedicated servers
03Lightweight ONNX Runtime inference option alongside optional GPU-backed PyTorch embeddings
04A/B recall mode for directly comparing and logging vector-only versus graph-augmented queries
05Per-turn state hook for tracking assistant turns, tool usage, and agent decisions with zero LLM overhead
0614 GitHub stars

ユースケース

01Implementing an automated search and recall layer for personal LLM-maintained Markdown wikis
02Supplying coding agents in Claude Code or Cursor with long-term associative memory over personal Obsidian vaults
03Evaluating the retrieval quality difference between pure vector search and 1-hop graph link traversal