Everything below is written for machines: what Bourdon is, how to integrate, what the guarantees are. Copy it into your agent’s context, or point the agent at bourdon.ai/llms.txt, the same content, served at the standard path agents check automatically.
# Bourdon > Bourdon is a recognition-first memory runtime for AI agents, built by RADLAB LLC. Unlike frameworks that optimize retrieval accuracy, Bourdon optimizes the time-shape of memory: recognition fires first (before details arrive), hydration runs in parallel, and archive descent happens only when needed. It provides one federated memory shared across the AI agents you already use (Claude Code, Claude Desktop, Codex, Cursor, GitHub Copilot, Cascade, Hermes, Devin, Lovable, and OpenClaw). Its sibling Positif (https://positif.ai) is the routing half of the same fleet OS. Key facts: - Status: pre-alpha. Python package 0.22.0 (2026-09-17); npm 0.18.0. Built in public. What pre-alpha means, which contracts are stable, and the CI gates: https://bourdon.ai/docs/status - Install: `pip install bourdon` (Python 3.10+), `brew install bourdonai/bourdon/bourdon`, or `docker run -d -p 7500:7500 -v bourdon-data:/data bourdonai/bourdon` - Core model: three-stage memory: (1) Recognition: near-zero-latency acknowledgment of who and what is present; (2) Hydration: parallel loading of relevant context while responding; (3) Archive descent: deep search only when the first two stages are insufficient. - Federation: participants read each agent's native store and publish a visibility-filtered L5 manifest (plain YAML, published JSON Schema); the L6 store indexes every manifest and serves them over MCP. Two servers can federate at depth one over HTTPS with bearer tokens. Mechanics: https://bourdon.ai/docs/federation - Memory format: https://bourdon.ai/docs/memory-format ; schema: https://bourdon.ai/schema/L5_manifest_v0.1.json - Security: loopback-only default bind, per-member tokens (hashed, shown once), trust tiers (quarantined writes are staged), append-only audit log, 421 on undeclared Host. https://bourdon.ai/docs/security - Evidence: dated field tests, each with method and limits: recognition latency 0 ms vs ~406 ms retrieval baseline (2026-04-19); recognition across a brand-new account (2026-05-15); cross-machine fallback memory expanded 588 B to 34,482 B (2026-05-26). https://bourdon.ai/field-tests - Licensing: Apache-2.0 for the CLI, the L5 wire format, the spec, and conformance fixtures; the engine is source-available under BUSL-1.1 and converts to Apache-2.0 four years after each release. Commercial licensing: [email protected]. - Falsifiable positioning: the project documents how it would know it is wrong (see the findings journal). ## Docs - [Docs overview](https://bourdon.ai/docs): the map, plus hosted and enterprise tiers - [Quickstart](https://bourdon.ai/docs/quickstart): install, setup, doctor, connect an agent over MCP, see recognition fire - [How federation works](https://bourdon.ai/docs/federation): participants, manifests, the L6 store, peers, drift - [The memory format](https://bourdon.ai/docs/memory-format): every L5 field, visibility, a complete example - [API reference](https://bourdon.ai/docs/api): MCP server transports and auth, the command line, stable contracts - [MCP tools](https://bourdon.ai/docs/mcp-tools): every tool and resource with signatures, generated from source - [Self-host](https://bourdon.ai/docs/self-host): stdio, Docker, always-on, claude.ai and ChatGPT via OAuth, two-instance federation - [Security model](https://bourdon.ai/docs/security): threat model, tiers, tokens, staging, audit - [Status and stability](https://bourdon.ai/docs/status): what pre-alpha means, release history, CI gates - [Thesis](https://bourdon.ai/thesis): why recognition-first, in full - [Related work](https://gitlab.com/bourdonai/bourdon/-/blob/main/spec/RELATED_WORK.md): how Bourdon differs from retrieval-first memory frameworks - [Findings journal](https://gitlab.com/bourdonai/bourdon/-/blob/main/spec/FINDINGS_JOURNAL.md): running log of field-test results ## Integration notes for agents - Connect over stdio with `bourdon serve` (Claude Code: `claude mcp add bourdon -- bourdon serve`), or over HTTP at `/mcp` on port 7500 with `Authorization: Bearer <token>`. - At turn start call `prepare_recognition_context`; it is bounded in time and never waits on retrieval. Call `get_deeper_context` afterwards if you need more. - Prefer `commit_to_federation` for durable facts and decisions; do not commit small talk. Commit under your own self-authoring `agent_id`, not under a reader-backed one such as `claude-code` or `codex` (those are regenerated from their native stores). - Memory is shared across the fleet: assume other agents can see what you commit unless the row is `visibility: private`. - Adapters exist for Claude Code, Claude Desktop, Codex, Cursor, GitHub Copilot (CLI and VS Code), Cascade (Windsurf), Hermes Agent, Devin CLI, and Lovable (cloud, export-format store, `bourdon lovable export`); OpenClaw is quarantined-class. Ten surfaces; one vendor is enough, since Copilot CLI and Copilot VS Code do not share memory with each other today. New adapters are welcome: https://gitlab.com/bourdonai/bourdon/-/blob/main/docs/AUTHORING_A_PARTICIPANT.md ## Family - [Positif](https://positif.ai): sibling project, payload-blind routing and audit for agent fleets. Bourdon = what the fleet knows; Positif = what the fleet is allowed to do. ## Optional - [Source on GitLab](https://gitlab.com/bourdonai/bourdon) - [Releases](https://gitlab.com/bourdonai/bourdon/-/releases) - [Field tests](https://bourdon.ai/field-tests) - [Blog](https://bourdon.ai/blog)