The science of
mesh cognition.
Mesh Cognition is the architectural pattern for Mesh Intelligence — center-free inference and learning among sovereign nodes — formalized as an open specification. Its foundational papers and normative specification are published and citable here — the canonical record.
Mesh learning toward mesh intelligence
Intelligence lives in the organization of a coordinator-free mesh — in how sovereign agents admit, bound, connect, and give birth to one another — not in the agents themselves. Mesh learning is the process by which such an organization improves itself from experience while its nodes are held fixed: better admission, sharper boundaries, richer accreted ontology, timely node-birth.
The wager is that an organization of ordinary cognitive nodes, learning only how to admit and grow, becomes collectively capable in ways no node is — and that this organizational intelligence is measurable independently of the nodes. Wiring language models together is not it. Wiring is where the intelligence is not.
The analogy is exact enough to be useful. A neuron is dumb; intelligence is not in the neurons but in the connectome — in how signals are selectively gated between parts. A brain survives replacing neurons and does not survive scrambling the connectome. Mesh intelligence is connectome-level intelligence over a collective of cognitive agents: the mesh gets more capable by reorganizing, not by upgrading its nodes.
Collective capability is a property of organization — separable and measurable independent of the nodes. Evaluated on frozen models, admission structure alone produces measurable collective gain, and, reported as found, sometimes collective harm.
What is Mesh Intelligence?
Mesh Intelligence is the emergent competence to infer and learn that arises, with no center, when sovereign agents each select what is relevant from what their peers emit and integrate it into their own state — the intelligence located in the selective coupling itself, not in any agent and not in a coordinator. Each agent holds a private, evolving cognitive state and emits only lossy, typed (CAT7) projections of it — never the state itself, and never by negotiation.
Every receiver admits what it finds relevant on its own terms, through a two-level coupling engine: a content gate (SVAF) that decides, per field, which dimensions of an emitted observation to absorb; and a temporal substrate that governs how absorbed, irregularly-timed contributions integrate into the receiver's state — a substrate that must be liquid, because the inbound timing is exogenous and cannot be scheduled.
From these distributed acts of selection and integration alone — no coordinator, no shared model — distinct sovereign nodes with different local views can produce coherent collective inferences that no single member can derive alone, up to the joint optimum of their views. This identification-completeness result is proven in the linear–Gaussian regime. Admission mechanizes coherence (relevance, not echo or noise); it does not mechanize grounding — whether accumulated inference becomes genuine knowledge rather than confident fabrication remains an open frontier.
Four composable primitives
carried by one protocol — the Mesh Memory Protocol. Only Cognitive Memory Blocks ever cross the wire.
The unit
Every agent emits a Cognitive Memory Block — a fixed seven-field, typed projection of its state. Lossy by design: the projection travels, the cognition does not.
The gate
Each receiver evaluates an incoming block field by field against its own role-indexed anchors — admitted, guarded, redundant, rejected, or silent. Admission is the receiver’s alone: no coordinator, no negotiation.
The substrate
Admitted fields fold into the receiver’s evolving state through a continuous-time (liquid) substrate where each neuron carries its own timescale — absorbing irregular arrivals rather than scheduled rounds.
The memory
A receiver stores its own evaluated understanding of what it admitted, content-hash-linked back to the source. Nothing echoed verbatim; every claim traceable. Lineage records provenance; it does not establish truth.
Two invariants make it center-free: hidden state never crosses the wire — sovereignty is structural, not policy — and receivers remix rather than echo, so there is no shared global state to coordinate.
Sovereign, distinct, and center-free
| Dimension | Mesh Intelligence | Swarm | Federated | Orchestration |
|---|---|---|---|---|
| Exchanged | Typed observations (CMBs) | Simple signals / stigmergy | Model gradients / weights | Messages, tool calls, hand-offs |
| Coordinator | None — receiver-autonomous | None, but trivial agents | Central aggregator | Orchestrator / graph |
| Each node holds | Private, distinct, evolving state | Minimal, identical state | A copy of one shared model | Role-prompted LLM, often stateless |
| Internal state | Never crosses the wire | n/a | Gradients can leak | Fully shared |
| Timing | Asynchronous, exogenous | Continuous local | Synchronous rounds | Turn-based |
| End state | Distinct sovereign states, coupled | Collective behaviour | One converged global model | Task done; nothing learned |
Swarm couples simple, identical agents; federated learning averages many copies into one model via a coordinator; orchestration routes tasks between agents at inference. Mesh Intelligence keeps agents sovereign and distinct, shares no weights and no raw signal, and supplies the missing layer — what each receiver does with an observation. It sits above tool- and task-level agent protocols, not against them. MMP is envelope-agnostic: SLIM, A2A, MCP or other transports can carry the blocks.
Pre-registered, then measured
Six real sensor networks, five domains. Every prediction — and its falsifying threshold — written down before the run.
No coordinator needed to reach the ceiling
Agents choosing their own peers beat a coordinator assigning one neighbourhood to everyone. Which peers help depends on the receiving agent’s own model, so no single assignment can be right for agents that differ.
Purpose changes what an agent admits
From identical messages under identical budgets, a receiver’s purpose changed which fields it took in — five times the variation the threshold required. This is the case for judging admission field by field rather than message by message.
Local selection reaches its own optimum
Free to take any number of peers from none to five, and scored against the best subset it could have chosen, receiver-local selection lost nothing at all on two domains.
The same result on the nonlinear river domain is not established
The median agent took a single peer and a quarter took none, but capture fell short of the bar set in advance. Reported because it was registered, not because it helps.
Per-field admission beats whole-message admission on the deployed corpus
Per-field won in every cell, but by a margin below the one we said in advance would count. Winning everywhere by too little is not the same as winning.
Two of five pre-registered predictions did not pass — reported as found, not as hoped. The mechanisms were evaluated across six real sensor networks spanning five domains with substantially different dependency structures, including one (river basins) with known physical upstream relationships. They do not, by themselves, establish equivalent properties for human or agent cognition, which remains the subject of the broader Mesh Cognition research programme.
Open, not established
Current research is examining when receiver-local policies preserve collective diversity and sufficiency, and when locally rational admission instead produces concentration or collapse. Operational sovereignty under heterogeneous utilities, cost-derived abstention, and coordinator-free collective sufficiency are open, not established. The informed-coordinator comparison, run under honest local supervision, is owed.
Foundational papers
Normative spec at meshcognition.org/spec/mmp — CC BY 4.0, v1.1.0 citable.