Graph Memory exposes 70 MCP tools, a REST API, and a WebSocket event stream from a single Node.js process. This post breaks down the architecture that makes it work: Graphology for storage, tree-sitter for AST parsing, serial queues for mutation safety, and hybrid search for retrieval.
Graph Memory generates vector embeddings for every node in every graph — doc chunks, code symbols, files, notes, tasks, skills. We run these embeddings locally using ONNX Runtime, not through an API like OpenAI's. This was a deliberate choice with real trade-offs.
We build Graph Memory with Graph Memory. Not as a marketing exercise — it's genuinely the fastest way for us to work. Here's a concrete example from a recent development session where we shipped six features in one sitting.