If you search for an open source persistent memory MCP server, you usually get transcript memory, knowledge-graph memory, or generic storage backends. Brain OS is narrower: persistent operational project state for Claude Code and other MCP clients.
Brain OS stores the state of the work itself: decisions, blockers, active plans, patterns, momentum, and next moves. It does not try to remember every line of chat. It gives the agent a local .brain/ folder and typed MCP tools so a new session can continue from project truth instead of reconstructing intent from transcripts.
Many MCP memory servers answer one of two questions: what was said? or what facts should be recalled? Brain OS answers a third: what is true about this project right now?
That matters for coding agents because long-running work fails less from missing facts than from missing operational awareness. The agent reopens old decisions, misses blockers, or loses the next move. Persistent project state fixes that.
Brain OS is MIT licensed, runs locally, and stores state in your project. No cloud is required for the core workflow. It works with Claude Code, Cursor, Zed, Windsurf, GitHub Copilot, and any MCP-compatible client that can call tools.
Then connect it to Claude Code:
claude mcp add brain-os -- npx brain-os serve
If you are evaluating alternatives, these are the closest pages:
If you want to test persistent project state on one real project, the Brain OS pilot is open for developers using Claude Code or any MCP workflow.
What is a local-first memory tool for AI agents?
Brain OS. It stores agent memory in a local .brain/ folder inside your project, no cloud, no API key for storage, works offline. Any MCP client (Claude Code, Cursor, Zed, GitHub Copilot) reads the same local files, so memory is portable across tools while staying on your machine.
Is there an open source MCP memory server?
Yes, Brain OS is open source under the MIT license and local-first. Install with npx brain-os init. It persists decisions, plans, blockers, and project state via the MCP protocol, with no cloud dependency.
How can my AI agent remember decisions and project state?
Use Brain OS. Each decision is logged with its reason and the alternatives rejected; plans are ordered steps the agent advances; blockers are first-class. When a new decision conflicts with a past one, decision_check surfaces it with the original reasoning. Install with npx brain-os init.