FIG.01 · MESH COGNITION
Mesh Intelligence on an open protocol. Pioneered by SYM.BOT.
FIG.02 · DEFINITION
Mesh Cognition is an architectural pattern for Mesh Intelligence in which each specialised agent emits semantically-typed projections of its own cognitive state — never the state itself — and each receiver admits them field by field, on its own terms, into its own sovereign store. Learned state stays per-node and is never merged.
Independent human, model, and process nodes, each emitting typed projections of its own local state (for example: focus, mood, intent, commitment) over a protocol. Each receiver decides locally what to admit, remix, and act on.
No merged state. No coordinating authority. Closer to how a research lab thinks: people overhear, take what's relevant, and the group converges without anyone in charge.
MMP carries typed Cognitive Memory Blocks. SVAF decides admission. Content-hash lineage keeps the trail reconstructible.
Six properties distinguish the pattern from centrally orchestrated multi-agent systems.
Each receiver admits or declines incoming observations field by field, never whole-message — each applying its own admission policy to each typed semantic field. The mechanism is specified as Symbolic-Vector Attention Fusion (SVAF); see MMP §9 for the definition.
What MMP carries between cognition nodes is a semantically-typed projection of cognitive state — a CAT7 Cognitive Memory Block, not a free-form message or tool call. Its fields come from the CAT7 vocabulary (focus, issue, intent, motivation, commitment, perspective, mood), so observations arrive typed — and per-field admission has something principled to decide over.
A receiver remixes what it admits rather than echoing it — so there is no shared global state to coordinate. A derived block retains content-hash lineage to its source observations through parent and ancestor references, making its cited provenance reconstructible across sovereign stores: its lineage identifies the source observations it cites.
By construction, no cross-node state merging. This is not a policy — it is an architectural property of the spec. Each node has one sovereign state boundary: learned state, when a node has it, remains local and is not merged across nodes by the protocol. A node may be a user, service, team, or org.
The protocol never requires learned node state to cross the wire. The open protocol enables interoperability without requiring a cloud or central server. No cloud is required by the spec.
Many multi-agent frameworks use an orchestrator or configured graph to select which agent acts. This pattern requires no central assigner: each node decides locally whether to admit an input, act on it, and contribute an observation. Participation is therefore self-selected — coordination is an emergent product of receiver-local decisions, not the output of a conductor.
admission · typed projections · remix · sovereignty · no required cloud · self-selection
FIG.03 · MOTIVATION
Biological intelligence is the structural precedent. The human brain has no single neuron that acts as a system-wide controller. Its tens of billions of neurons receive many inputs, commonly at synapses whose effects vary in strength, integrate them through changing membrane and biochemical dynamics, and signal onward through action potentials and synaptic release. No single neuron directs the whole; perception and thought depend on activity distributed across interacting circuits — though how those processes produce cognition remains an open scientific problem. Mesh Cognition takes limited local units, connection-specific influence, partial signalling, and the absence of a single controller as structural inspiration — not biological equivalence — for collective intelligence among agents. In the architecture, the cognition node is the sovereign unit: a participant behind its own admission boundary.
The idea has a long lineage. Psychology, cognitive science, and philosophy hold several influential accounts of distributed cognition: Minsky's society of mind, in which simple, individually mindless processes compose intelligent activity; distributed-cognition accounts spanning people and artefacts, and the extended-mind argument that some external resources constitute parts of a cognitive process; research on transactive memory, where groups organise access to knowledge held across their members. Across these distinct traditions, a recurring motif is coordination among limited, differently situated contributors, with no one contributor holding the whole.
Many agentic systems use the opposite pattern. One model, one context, one perspective — and coordination between agents routed through a conductor. Some hard problems benefit when specialists work in parallel and build on one another's observations; for those, model capability is not the only constraint — coordination bandwidth among participants also matters.
Centralized coordination doesn't scale to autonomy. If every agent has to ask a controller, the controller is the ceiling. Receiver-autonomous admission lets the mesh grow without re-introducing a master: no central admission controller is required.
Sovereignty unlocks institutions. Institutions require local control of learned state and auditable exchange boundaries. They can emit typed projections into a protocol while learned state remains local and cited lineage remains reconstructible.
Collective cognition adds infrastructure requirements on top of model requirements.
FIG.04 · POSITIONING
Position statement Mesh Cognition is not a product, an orchestrator, or a network primitive. It is the architectural pattern for Mesh Intelligence with per-node sovereignty, formalized as an open standard: nodes emit typed projections of their own cognitive state, and each receiver decides what to admit and act on.
The emergent competence to infer and learn, with no center, among sovereign agents. research ↗
The architectural pattern for it; its unit is the cognition node. The subject of this site. the node ↗
The open standard that formalizes the pattern — typed blocks, per-field admission, remix with lineage. CC BY 4.0. spec ↗
SYM is the maintained open reference implementation; independent emitters and integrations are welcome. list ↗
It is often mistaken for its neighbours — orchestration frameworks, federated learning, swarm. The distinction is in what crosses between agents, and whether coordination is receiver-autonomous or directed by a coordinator: see how it differs.
FIG.05 · FOUNDATIONS
Mesh Cognition is grounded in five research preprints and one open specification.
"On the Necessity of a Liquid Substrate for Mesh Intelligence"
The substrate condition: any fixed-weight agent folding peers' projections online must meet two necessary conditions — an adaptive timescale, and a dependence on the elapsed gap between irregular arrivals that no gap-blind network recovers at any width or depth. Necessary, not sufficient, for a fixed-weight receiver that integrates irregularly arriving peer projections online.
"Mesh Inference: A Formal Model of Collective Inference Without a Center"
The proof layer: independent agents, exchanging only admitted typed observations with no shared weights or hidden state, derive a conclusion none holds alone. One admission/emission policy governs three properties — convergence, identification-completeness, and observation-only confidentiality.
"Mesh Memory Protocol: Semantic Infrastructure for Multi-Agent LLM Systems"
The protocol paper defining typed observations, receiver-local per-field admission, content-hash lineage, and remix graphs; the open specification supplies the normative wire format.
"MeloTune: On-Device Arousal Learning and Peer-to-Peer Mood Coupling for Proactive Music Curation"
Application paper: on-device emotion-aware curation through peer-mesh. Per-listener arousal adjustment learned from behavioral signals, integrated into a continuous-time curation pipeline deployed on iOS.
"Symbolic-Vector Attention Fusion for Collective Intelligence"
The per-field admission mechanism: how an agent accepts or rejects incoming fields with role-indexed anchors and four-class admission outcomes (aligned / guarded / redundant / rejected).
Open protocol specification
The canonical wire-protocol specification at meshcognition.org/spec/mmp, licensed CC BY 4.0. Editor: Hongwei Xu.
FIG.06 · REFERENCE IMPLEMENTATIONS
The open MMP wire contract supports Class 1 emitters; SYM is the open reference runtime for cognition-node behavior. Production consumer apps embed the emitter SDKs.
Open specification at meshcognition.org/spec/mmp under CC BY 4.0; the wire contract — Class 1 emission is the third-party surface (§17.1).
spec ↗Node.js mesh runtime; npm-installable; CLI for joining AI copilots into a personal mesh.
npm ↗ github ↗Autonomous-LLM-peer runtime for dedicated agents that wake on incoming CMBs.
npm ↗ github ↗Bridge that pairs Claude Code (and other AI coding agents) into the mesh as participating peers.
npm ↗ github ↗Swift emitter SDK for Apple platforms; embedded in the production iOS apps. Speaks an earlier MMP revision; current-scheme Class 1 conformance in progress.
github ↗Build on Mesh Cognition? Open a discussion — third-party emitters and integrations welcome.