The wire protocol for mesh cognition — collective intelligence with no center.
The Mesh Memory Protocol specification is published at meshcognition.org/spec/mmp.
- Browse the spec by section: meshcognition.org/spec/mmp
- Single-page Markdown: meshcognition.org/spec/mmp-v1.0.md (in this repo)
- Single-page HTML: meshcognition.org/spec/mmp-v1.0.html (in this repo)
- Governance + RFC process: meshcognition.org/governance
spec/— the single-page specification (Markdown + HTML), mirrored from the canonical site. The site is the source of truth; these are kept in sync for git history, offline reading, and GitHub rendering.extensions/— community extension drafts, edited via pull request. Promoted to core MMP per the §16.5 lifecycle.- Issues — public RFC discussion forum for spec changes. File proposals here with
[spec-rfc]in the title. - License: CC BY 4.0 (specification text). Reference implementations are licensed Apache 2.0.
| Language | Project | Scope |
|---|---|---|
| Node.js | sym-bot/sym |
Full L0–L7 |
| Swift | sym-bot/sym-swift |
Full L0–L7, Apple platforms |
| Python | sym-bot/mesh-cognition (pypi) |
Coupling kernel: L4 (per-field admission) + L6 (Cognitive State) for CfC neural networks |
| Node.js | sym-bot/xmesh-agent |
Autonomous agent runtime — Anthropic / OpenAI / Ollama on the MMP wire |
| Node.js | sym-bot/sym-mesh-channel |
Real-time Claude-to-Claude mesh (Claude Code plugin) — first non-Anthropic Channels implementation on MMP |
MMP is one layer of a characterised stack — which to admit (SVAF) · whether a center-free collective can recover the answer (Mesh Inference) · how each agent tracks an evolving latent in time (Liquid Necessity) — carried by the protocol and shown in a deployed reference. The canonical record is meshcognition.org/research:
| Paper | arXiv |
|---|---|
| On the Necessity of a Liquid Substrate for Mesh Intelligence — for any fixed-weight agent folding peers' projections online, an adaptive timescale and gap-awareness are necessary (necessary, not sufficient) | 2606.28413 |
| Mesh Inference — A Formal Model of Collective Inference Without a Center (convergence, identification-completeness, observation-only confidentiality) | 2606.19537 |
| MMP — Mesh Memory Protocol: Semantic Infrastructure for Multi-Agent LLM Systems | 2604.19540 |
| MeloTune — On-Device Arousal Learning and Peer-to-Peer Mood Coupling (first deployed reference) | 2604.10815 |
| SVAF — Symbolic-Vector Attention Fusion for Collective Intelligence (the per-field admission gate, L4) | 2604.03955 |
If you use MMP in your research, cite the protocol paper:
Xu, H. (2026). Mesh Memory Protocol: Semantic Infrastructure for Multi-Agent LLM Systems. arXiv:2604.19540.