Graymatter

GrayMatter is a persistent memory layer for AI coding agents. Agents are stateless by default, so each run feeds the whole conversation history back in and the token cost grows with it. GrayMatter stores facts locally, recalls them by meaning and injects only the handful that matter to the task at hand. It is built for developers running agents in editors and terminals. It ships as a single static binary of around ten megabytes with no Docker, database, configuration file or cloud account, and works as a general purpose MCP server with Claude Code, Cursor, Codex, OpenCode, Antigravity and other MCP clients, or as a plain Go library. The project reports a ninety percent reduction in context tokens by injecting the top eight relevant facts instead of the full history, and says every published figure is checked in continuous integration. A self building knowledge graph records typed entities and co mention edges from ordinary use and exports to Obsidian. A reflect tool lets the agent add, update, forget and link its own memories, a context sync command projects top facts into agent instruction files within an explicit token budget, and memories fade on a deterministic thirty day half life using tombstones rather than deletions. The project is free, runs offline and requires no account, and there is no pricing page or paid tier. Installation is through Homebrew on macOS and Linux, with Scoop, Nix, prebuilt binaries and a Go install route also documented, followed by commands that wire up the MCP configuration and verify the setup. The authors are explicit about what it is not: not a framework, not a hosted service and not a knowledge base interface.