The Open Foundation

Mesh Memory Protocol (MMP).

The open protocol for Organizational Cognition — how sovereign AI agents exchange cognitive influence without a coordinator, without merged state, and without surrendering their own reasoning.

  1. Organizational Cognition
  2. Mesh Cognition
  3. MMP
  4. xMesh

Why MMP exists

Organizational judgment cannot be coordinated through a center: a decision-maker that can be overwritten by its inputs isn’t deciding anything, and a shared store everyone must accept is a center by another name. For independent decision-makers to influence one another selectively — the mechanism Organizational Cognition depends on — the exchange itself needs rules: what crosses the wire, who decides what is admitted, and how provenance survives. MMP is those rules, published openly so no one — including us — owns the exchange.

What the protocol specifies

01
Typed projections

An agent emits semantically-typed projections of its cognitive state — seven fields (focus, issue, intent, motivation, commitment, perspective, mood) — never the state itself.

02
Per-field admission

The receiver decides locally what to accept, field by field, under its own admission policy. No sender and no coordinator can force admission of anything.

03
Remix with lineage

Admitted observations are remixed into the receiver’s own memory — never echoed — with content-hash lineage back to the source observation.

04
Sovereign per-node state

Learned state, when a node has it, stays local. The protocol never merges state across nodes and never requires state to cross the network.

05
No required cloud

Nodes couple directly. There is no mandated central service, shared store, or platform dependency in the protocol.

06
Self-selection, not assignment

Nodes decide for themselves whether to participate and what to admit. Coordination emerges from distributed, receiver-local decisions.

How it differs

  • Orchestration frameworks route work through a coordinating authority. MMP has no coordinator — influence moves through receiver-local admission decisions.
  • Federated learning pools model gradients or weights toward one shared model — a center by another name. MMP shares no weights and builds no shared model.
  • Swarm systems exchange simple signals. MMP exchanges typed cognitive observations that receivers evaluate, admit and remix with provenance.

Open governance

  • The specification — published openly (CC BY 4.0) at meshcognition.org, the standards and research home. The current version lives there, not here.
  • SYM — the Apache-2.0 reference runtime and CLI.
  • sym-mesh-channel — the Claude Code channel surface built on SYM.
  • Build on it — connect Codex, Claude Code and other agent surfaces.
Read the Specification Build on it — Developers Explore Organizational Cognition