FIG.01 · MESH COGNITION
A new architectural pattern where specialised agents share semantically-typed observations through per-field admission, with sovereign per-user model state. Pioneered by SYM.BOT.
FIG.02 · DEFINITION
Mesh Cognition is a distributed-intelligence architectural pattern in which specialised agents share semantically-typed observations through a wire protocol with per-field admission, while each agent maintains sovereign learned state that is per-user and not merged across users.
Many independent cognitions — humans, models, processes — sharing partial state (focus, mood, intent, commitment) over a protocol. Each receiver decides locally what to admit, remix, and act on.
Not a hive mind. Not federation. Closer to how a research lab thinks: people overhear, take what's relevant, and the group converges without anyone in charge.
MMP carries the thoughts. SVAF decides admission. Content-hash lineage keeps the trail reconstructible.
The pattern is defined by five distinguishing properties.
Agents accept or reject incoming observations field-by-field, never whole-message. Each agent carries learned admission weights per typed semantic field. The mechanism is specified as Symbolic-Vector Attention Fusion (SVAF) — see the SVAF paper for the formal definition and the MMP §4 specification for the wire-protocol form.
Every claim is traceable to its source observation. Cognitive Memory Blocks (CMBs) carry content-hash keys with parent and ancestor references, stitching the collective memory into a verifiable remix graph. A receiver can always answer: which observations gave rise to this state?
By construction, no cross-user state merging. Each user's learned model is structurally isolated. This is not a policy — it is an architectural property of the spec. Sovereignty here means one model instance per user, not merged across users by the protocol. See the MeloTune paper for the deployed reference.
Temporal trajectories integrate through closed-form continuous-time networks, not discrete-step. This permits irregular event timing natively (no resampling, no event-binning) and produces analytic state updates suitable for sub-millisecond inference on consumer hardware.
Privacy-by-architecture: learned state stays on the user's device. The open protocol enables interop without server dependency. No cloud is required by the spec.
admission · lineage · sovereignty
continuous-time · on-device
FIG.03 · MOTIVATION
Single-agent cognition is the bottleneck. One model, one context, one perspective. Hard problems get better when specialists think in parallel and share state. The constraint is no longer model capability — it is coordination bandwidth between cognitions.
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 — same reason TCP/IP beat circuit-switching.
Sovereignty unlocks institutions. Hospitals, labs, banks cannot pour data into someone else's brain. They can share observations into a protocol where state stays theirs and lineage is auditable.
Cognition is becoming an infrastructure problem, not a model problem.
FIG.04 · POSITIONING
Position statement Mesh Cognition is not a product, an orchestrator, or a network primitive. It is the architectural pattern for distributed intelligence with per-user sovereignty, formalized as an open protocol.
FIG.05 · FOUNDATIONS
Mesh Cognition is grounded in three peer-reviewed / preprint research papers and one open specification.
"Mesh Memory Protocol: Semantic Infrastructure for Multi-Agent LLM Systems"
The eight-layer protocol that defines typed observations + per-field admission + content-hash lineage + remix graph. The wire substrate every Mesh Cognition implementation speaks.
"Symbolic-Vector Attention Fusion for Collective Intelligence"
The per-field admission mechanism: how each agent learns to accept or reject incoming fields with role-indexed anchors and four-class admission outcomes (aligned / guarded / redundant / rejected).
"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.
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
Reference implementations across runtimes and Apple platforms, with production consumer apps deployed on top.
Open specification at meshcognition.org/spec/mmp under CC-BY-4.0; the substrate every reference implementation below speaks.
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 ↗Python coupling kernel; per-field admission and state blending for CfC models. Pure Python, zero external dependencies.
pypi ↗ github ↗MeloTune (iOS) + MeloMove (iOS) on Mesh Cognition; MeloMotion (web, MMP integration in flight) demonstrates cross-domain value (mood ↔ activity ↔ creative state).
Build on Mesh Cognition? Open a discussion — third-party implementations welcome.
INDEX · ENTRANCE LINKS
Version-tag schedule, change-log per release, editor responsibilities. Coming soon.
MMP, SVAF, MeloTune; forthcoming preprints; adjacent work. Coming soon.
Independent libraries, ports, runtimes beyond the SYM.BOT stack. Coming soon.
Discussion home for the open protocol. Coming soon.