- A brain for AI is an external, persistent store your AI tools query over MCP, so knowledge survives sessions and is shared across every agent you run.
- Setup is fast: create a brain, connect one agent, seed a few facts, and run the fresh-session recall test. Ten minutes covers it.
- Seed decisions and durable facts, not transcripts; one clear fact retrieves better than a pasted wall of notes.
- The recall test is the whole point: tell it something, start a completely new session, ask for it back. If that works, you have a memory.
- This guide is the personal and agent-first path. Rolling a brain out to a company is a different job, covered in our company AI brain guide.
To set up a brain for AI: create a governed brain (free, one command: npx @aivm/brain init), connect your first agent (a plugin for Claude Code, one-command installs for Cursor and Claude Desktop, an MCP config block for anything else), seed it with a few durable facts, then prove recall from a fresh session. From then on, capture happens as you work.
Before you start: what you are actually building
This guide sets up a personal brain: your knowledge, your agents, ten minutes. (Rolling out a brain for a whole company, with connectors into Slack and Drive and a security review at the end, is a different project; that walkthrough is our company AI brain guide.)
The thing you are building is simple to say: one store that remembers what you and your AI tools learn, and answers when any of them asks. The reason it has to be external infrastructure, not a notes file, is reach. A CLAUDE.md helps exactly one tool in exactly one repo. A brain over MCP serves every MCP-capable agent you run today and the ones you will run next year.
Step 1: create the brain
Run npx @aivm/brain init, or sign up at brain.aivm.io. Free tier, no sales call. You get an isolated tenant (per-tenant Postgres isolation, not a shared index), a dashboard, and a key mint for connecting agents. This is also where governance lives when you need it later: roles, redaction, the access ledger.
Step 2: connect your first agent
Pick the agent you actually use most; one honest connection beats five stubs. Claude Code gets the deepest integration: a plugin that hooks the session lifecycle, so capture and recall happen automatically while you work. Cursor and Claude Desktop take a one-command install that patches their MCP config. Codex, Hermes Agent, OpenClaw, and any other MCP-capable client take a standard server block with your key in it.
Each agent has its own setup page with the exact steps under our brains-for-agents hub. All of them end the same way: restart the agent, and it can now read and write your brain.
Step 3: seed it with what you know
Give it something worth remembering. Upload a few documents that describe your world (a project brief, your resume, the architecture doc), then capture five durable facts by just telling your connected agent to remember them: what you are building, what stack you use, what you decided last week and why.
Quality rule: decisions with reasons beat raw material. 'We ship on Render, autodeploy from master, never push on Fridays' is a great memory. A meeting transcript is compost.
Step 4: run the recall test
This is the step people skip and should not. End your session completely. Open a fresh one (new terminal, new chat), and ask about a fact you seeded: 'what stack do we deploy on?' If the answer comes back correct, from your brain rather than the model's guess, you have working AI memory. If it does not, the setup page for your agent has the troubleshooting list (the usual suspects: the agent was not restarted, or the key never landed).
Step 5: let it compound
From here the habit is small: when something durable happens, make sure it lands in the brain. With the Claude Code plugin that is mostly automatic; elsewhere it is one instruction ('save that to the brain'). Add your second agent whenever you like; it inherits everything the first one stored. That cross-agent inheritance is the compounding effect, and it is what a per-app memory feature can never give you.
When a second person joins the brain, read the governance guide first: permissions, redaction, and the audit trail are what make shared memory safe.