# DecisionNode

A locally stored, semantically searchable store of architectural decisions that an AI agent can query through MCP or a developer can browse through a command line and a graph-based web interface.

DecisionNode is a shared, structured memory store for architectural and project decisions, exposed to Claude Code, Cursor, Windsurf, Antigravity and other MCP clients. Rather than a markdown file kept in an assistant's context the whole time, a decision is stored as JSON, embedded as a vector using a Gemini embedding model, and retrieved later only when it is actually relevant to a query.

A decision can be added directly through the command line or by an AI calling a tool over MCP; either way, it is embedded and stored locally, and later retrieved by another MCP call that searches by meaning and ranks results by similarity, with nothing pre-injected into the system prompt. A local web interface offers three views of the same store: a force-directed graph of decisions and their similarity, a projection of the embeddings into two dimensions, and a searchable, filterable list.

Both the command line and the MCP server read and write the same local store.

Website: https://decisionnode.dev