Bourdon is live: one memory for every AI you use

The problem we kept living

We run a fleet of AI coding agents here: Claude Code, Codex, Cursor, Copilot, and a few of our own. Tell one of them something on Monday and the rest still don't know it Tuesday. Every tool keeps its own memory, in its own silo, on someone else's infrastructure. And when an agent does remember, it remembers like a filing cabinet: silence, retrieval, then a paragraph read back at you.

Human memory doesn't work that way. When you see a familiar face you don't freeze to search; you recognize first and the details arrive while you're already talking. Bourdon is that behavior, built as infrastructure: recognition fires in milliseconds from a hot cache while deeper context hydrates in the background, and every agent reads and writes the same federated memory.

What makes it different

Your memory is a file, not an account. A directory of plain YAML on your own disk, in a published schema, served over MCP. You can grep it, put it in git, sync it between machines, or walk away with it. Sixteen agent integrations ship in the box, and several of them, we're fond of noting, wrote their own connections: Copilot and Windsurf's Cascade each authored their own Bourdon participant, and our in-house agent Hermes wired up its own too. The tools that built it are the tools it gives a shared memory to.

Nothing leaves your machine unless you say so. Peer federation is strictly opt-in: no peers load without an explicit flag, and the receive path quarantines and validates everything before it touches your library.

The two days before launch

This launch ships version 0.13.0, and the number carries some scar tissue. Last weekend we ran a full three-star review of the codebase (the exhaustive kind, with independent adversarial review lanes), and it came back with a stop-ship verdict: five release-blocking findings and thirty-four must-fix issues. Two days later, all of them were closed: quarantined sync, zero-egress federation defaults, a protected release chain publishing through OIDC with no long-lived credentials, resource budgets, schema validation, the lot. The verify gate even caught its own release pipeline misconfigured on the first attempt and refused to publish, which is exactly what it's for. The full release record ships in the repository, because a memory tool you're trusting with your working context should show its work.

Try it

Free, self-hosted, yours:

pip install bourdon && bourdon setup
# or the MCP server via npm:
npx @getbourdon/mcp-server --library ~/agent-library

Start at bourdon.ai, read the source at gitlab.com/bourdonai/bourdon, and if the recognition-first idea resonates (or you think we're wrong about something), we'd genuinely like to hear it on the Product Hunt thread today.