divyanshailani/graph-memory

Graph Memory is a local database tool built to give AI coding agents, such as Claude, Codex, and other assistants, a long term memory of a codebase's structure. It stores information as a graph of nodes and relationships in a local SQLite file, and it connects to AI agents through the Model Context Protocol, a standard way for agents to read and write structured project information as they work. One of its core features scans a real codebase and turns it into a graph automatically, using a parsing technique that understands the actual structure of the code rather than just its text, and it supports Python, TypeScript, JavaScript, Go, and Rust. It can also generate short summaries of each part of the code using an external AI service, and it supports exporting the graph as an HTML file so a person can visually inspect what the agent has learned about the project. A notable design choice is trust scoring, which separates facts that are known for certain, like something confirmed by parsing the actual code, from assumptions an AI agent generated on its own, helping keep unreliable guesses distinct from verified structure. The project also defines specific categories of information nodes can hold, separating hard facts, higher level knowledge like design decisions, and records of completed work such as how a particular bug was fixed. To use it, a person installs it with pip, configures it as an MCP server in their AI tool's settings, then runs a command to scan a codebase and, optionally, an environment variable pointing to an API key to generate the automatic summaries. The project describes itself as inspired by an existing open knowledge format for AI agents, but built with a real database and code aware parsing instead of static text files, aiming to reduce mistakes that come from AI generated information alone.