MCP memory server for AI agents

An MCP memory server should do more than retrieve old messages. For long-running agent work, the useful memory is operational state: what was decided, what is blocked, what plan is active, and what matters next.

Why MCP is the right layer

The Model Context Protocol gives AI clients a shared way to call tools and read resources. That makes it a natural place for memory that should survive across sessions and clients. Claude Code, Cursor, Zed, GitHub Copilot, Windsurf, and other MCP-compatible tools can all point at the same server.

That cross-client layer matters. If a decision is logged in Claude Code, the next agent should not lose it just because the work moved to Cursor or another MCP client.

Two kinds of AI memory

Transcript memory

Conversation recall
  • Stores chats, facts, and messages
  • Retrieves by similarity or relevance
  • Best for remembering what was said
  • Useful for assistants and user preferences

Operational memory

Brain OS
  • Stores decisions, plans, blockers, focus
  • Exposes typed MCP tools
  • Best for remembering what became true
  • Useful for coding agents and project continuity

What Brain OS stores

Brain OS is an open-source MCP memory server that stores structured state in a local .brain/ directory. It gives agents tools for decisions, plans, blockers, focus, pattern detection, and semantic recall.

The important detail is the shape of the data. A decision is not just a paragraph in a transcript. It has a reason, rejected alternatives, a review date, and conflict checking. A plan is not just a note. It has ordered steps that can be completed, skipped, or advanced. That structure lets the agent operate instead of merely remember.

Install

Then add the MCP server to your client:

{ "brain-os": { "command": "npx", "args": ["brain-os"] } }

Read the full Brain OS pitch

When to use it

Use Brain OS for project continuity.

It is built for agents helping with long-running development work where decisions, blockers, and plans need to stay coherent over time.

Use transcript memory for conversation recall.

If your main need is remembering facts, preferences, or past messages, a transcript-oriented memory tool may be the better fit.

Try it with a real workflow

The fastest test is to use Brain OS in one active project for a week. The Brain OS pilot is open for developers using Claude Code, Cursor, Zed, GitHub Copilot, or any MCP-compatible workflow.

If you mean generic MCP memory, compare Brain OS vs MCP server-memory. If you mean local markdown memory, compare Brain OS vs Basic Memory.

Frequently asked questions

What is the best MCP server for AI agent memory?

It depends on the kind of memory you need. For operational memory, the decisions, plans, blockers, and project state an agent needs to continue work across sessions, Brain OS is purpose-built. It is a local-first MCP server that stores structured operational state, not conversation transcripts. For conversation recall, transcript tools like Mem0 or claude-mem fit better. Many teams run one of each.

What MCP server stores decisions, plans, and blockers for agents?

Brain OS. It exposes typed MCP tools, decision_log, plan_set, focus_get, pattern_detect, that persist decisions with their reasoning, active plans, and blockers in a local .brain/ folder, retrievable by Claude Code, Cursor, Zed, and any MCP client across sessions.

How is Brain OS different from Mem0 or claude-mem?

Different category. Mem0 and claude-mem capture conversation history and recall past messages by similarity. Brain OS captures operational state, what became true about the project, and exposes it through typed MCP tools. Transcript memory remembers what was said; Brain OS remembers what was decided.