BrainMesh

The follow graph for second brains.

Personal AI memory layers are exploding: GBrain (26k+ stars), Obsidian vaults behind MCP servers, and a dozen other markdown-first stores. Every one of them is a silo. GBrain already knows how to publish public subsets, mount federated read-only sources, and serve a brain over MCP with OAuth. What nobody has built is the layer between strangers: identity, discovery, and trust.

BrainMesh is that layer. An open manifest spec plus a public registry so any brain can be found, previewed, and followed by any other brain, on any stack.

Think npm for brains: the protocol and registry are free and open forever, and anyone can build hosted services on top.

What you can do with it

1. Publish: describe your brain in a small brain.json manifest (identity, MCP endpoint, public topics, access terms) and submit it to the registry via PR.
2. Discover: browse or search the registry index for brains covering topics you care about.
3. Preview: read a brain's human-readable BRAIN.md and access terms before trusting a byte of it.
4. Follow: mount a brain's published public subset as a read-only federated source in your own memory layer, so your agent can search it like local knowledge.
5. Audit and unfollow: list every external brain wired into your agent, with provenance, and cut any of them loose in one command.

Architecture

flowchart LR
    subgraph you [Your machine]
        A[Your agent] --- B[(Your brain<br/>GBrain / Obsidian / other)]
    end
    subgraph alice [Alice's world]
        C[Public subset repo<br/>markdown on GitHub]
        D[Alice's MCP endpoint<br/>gbrain serve --http]
    end
    R[(BrainMesh registry<br/>manifests + index.json)]

    B -- 1. publish manifest via PR --> R
    A -- 2. search + preview --> R
    R -. points to .-> C
    R -. points to .-> D
    B -- 3. follow: clone + mount<br/>read-only source --> C
    A -- 4. optional live query<br/>scoped read token --> D

Loading

Two follow modes, both riding on machinery that already ships today:

• Repo mode (default): clone the brain's published public subset and register it as a read-only source in your own brain. Works offline, survives the publisher's laptop sleeping, and keeps you in control of what enters your context.
• Endpoint mode (optional): query the publisher's live MCP endpoint with a read-scoped token. Fresher, but requires their server to be up and requires you to trust their access terms.

Quickstart

Publish your brain

# 1. Copy the template and fill it in
cp spec/examples/example.brain.json my-handle.brain.json

# 2. Validate locally
node scripts/validate-manifests.mjs my-handle.brain.json

# 3. Fork this repo, drop the file in registry/brains/, open a PR
#    CI runs the same validator. Merge = listed.

GBrain users: the brain-network skillpack in skillpack/ automates all of this, including generating the public subset repo. See skillpack/brain-network/README.md.

Follow a brain (GBrain)

# Clone their public subset and mount it as a read-only source
git clone https://github.com/alice/brain-public ~/brain-network/alice
gbrain sources add bn-alice --path ~/brain-network/alice
gbrain sync --source bn-alice
gbrain query "what does alice know about agent memory?"

Read SECURITY.md before following anyone. Mounting a stranger's markdown into your agent's context is a supply chain decision, and BrainMesh treats it like one.

Follow a brain (Obsidian)

See docs/adapters/obsidian.md for exposing a vault through an Obsidian MCP server with a publish frontmatter flag, and for consuming published subsets into a vault.

Repo layout

Path	What it is
spec/brain-manifest.schema.json	JSON Schema for brain.json (the wire format)
spec/BRAIN.md.template	Human-readable companion doc every brain publishes
registry/	The v0.1 registry: brains/*.json submissions plus a generated index.json
scripts/validate-manifests.mjs	Dependency-free validator, run locally and in CI
skillpack/brain-network/	GBrain skillpack: publish, search, follow, unfollow/audit
docs/adapters/obsidian.md	Obsidian vault adapter guide
SECURITY.md	Threat model and mitigations for mounting third-party brains
ROADMAP.md	v0.2+: hosted registry, relay, verified identity, paid subscriptions

Design principles

1. System-agnostic: the manifest describes any MCP-fronted markdown store. GBrain and Obsidian are the first two adapters, not the spec.
2. Git is the transport: v0.1 needs no new servers. Manifests travel by PR, brains travel by clone. Boring on purpose.
3. Read-only by default: following grants your agent read access to their published subset, never write access anywhere.
4. Trust is explicit: nothing enters your agent's synthesis path until you have previewed the manifest, the terms, and (recommended) quarantined the first sync.
5. Open core: protocol and registry stay MIT. Monetization lives in optional hosted services on top. See ROADMAP.md.

Status

v0.1, pre-launch. The spec is a draft and breaking changes are expected until v1.0. Feedback via issues is very welcome, especially from GBrain, Obsidian MCP, and agent-harness maintainers.

License

MIT. See LICENSE.