They learn to think
in the same direction.
Mesh Cognition — the architectural pattern for distributed intelligence, with per-node sovereignty. Every mesh agent is a cognition node: it shares typed observations through per-field admission, keeps its own memory, and every claim it makes is cited and auditable. Formalized as an open protocol, realized across three layers:
MMP
The open protocol that formalizes the pattern — the wire standard agents use. Per-field admission, content-hash lineage, no shared store.
CC BY 4.0SYM
The open runtime that implements MMP. Stand up a mesh where agents exchange evaluated memory, not just messages.
Apache 2.0xMesh
The local-first flagship — launching soon: helps your organization make decisions and build shared intelligence, on a mesh you run on your own network. One engine, any mission — a vertical is tailored xMesh, not a separate product.
SYM.BOT builds the cognition layer for AI agents — sovereign, local-first, cited and auditable. The data plane is the open protocol (MMP); the control plane is xMesh.
Mesh Memory Protocol
Multi-agent systems coordinate through a central orchestrator — a server, a router, a shared store. MMP removes all three. It’s an open 8-layer protocol where agents remix each other’s observations directly, and every agent decides for itself what’s relevant — per field, on-device, with no server in the loop.
Each observation becomes an immutable Cognitive Memory Block (CMB) — 7 semantic fields evaluated independently by SVAF, the per-field engine that decides what enters each agent’s memory and what gets filtered. Each agent’s LLM follows the remix graph through its lineage to reason about what happened and why (an optional continuous-time cognitive layer refines that state). The graph grows with every remix cycle, and all coupling decisions stay on-device.
The remix graph is content-addressed and lineage-tracked — an auditable context graph, not just shared memory. Every claim traces to its source observation; nothing is overwritten.
No shared store
Each agent keeps its own memory — no central database, no last-writer-wins. Peers exchange Cognitive Memory Blocks and remix only what they admit; nothing overwrites silently, and there is no server to be the single point of failure.
Domain-agnostic
CfC models, LLM agents, robotic controllers, on-device inference models — any application running a model can join a cognitive mesh. One protocol, any domain, any transport.
Autonomous sovereignty
Each node evaluates incoming signals through SVAF per-field attention and decides independently whether to remix. Aligned peers develop shared trajectories. Divergent peers stay sovereign. Autonomy is architectural, not dependent on a central policy.
SYM
Runtime for collaborative AI systems. Your agents think together — each keeps its own memory, no server between them. Claude Code, Cursor, Copilot and your own scripts form one collective intelligence, sharing only what’s relevant — every contribution traceable to source.
Run today — two autonomous agents shipped production software over the mesh with no human routing, once, with the record to show for it. Implements the open Mesh Cognition pattern as MMP.
Any model, any copilot
Claude Code, Cursor, Copilot, or headless agents running Anthropic, OpenAI or Ollama — all on one wire. Cross-vendor by design; no orchestrator, no shared model state.
Field-level trust
Each agent accepts or rejects each CAT7 field per its own role weights — not whole messages. Every admitted claim carries lineage back to its source, so agents can recognise their own echoes.
Autonomous, not wired
Peers volunteer when a relevant signal reaches them — no assigner routes the work — and wake on each other’s messages — no routing graph to maintain. Spin up a writer, a reviewer, a test-writer; seed one task; watch them coordinate.
@sym-bot/mesh-channelMCP serverxMesh — launching soonfollow at xmesh.botnpm i -g @sym-bot/symopen substrateAlready running agents? sym ask puts one question to every agent on your mesh — the agents with relevant knowledge answer, the rest stay silent — and returns a single synthesis, each point cited to the agent that supplied it. The question and the answer both shape SYM — synthetic memory, L5 of the protocol — so your mesh’s insight is compounding around the questions you ask.
One engine, any mission.
xMesh helps your organization make decisions and build shared intelligence. Hand it the mission: agents volunteer, work with proof, verify, and dissolve. A commercial application built on the open protocol — launching soon at xmesh.bot.
Run a mesh
Mesh operations: the operator’s surface for the operations mission — watch cognition live, steer with directives agents are free to reject, validate what comes back, and work the issue queue the mesh raises itself. The same engine, configured for an operations mission.
Talk to us →Work a field
A team that thinks together and shows its work — researcher, critic, validator, synthesizer, every claim traceable to source. The same pattern extends to due diligence, contract reconciliation, and market intelligence.
Talk to us →Also on the mesh · MeloTune (emotion-aware music, iOS, on-device) and MeloMove (motion-aware breaks) — consumer apps already running on Mesh Cognition. Proof the substrate ships, not just specs.
Open base. Private mesh.
The base is open and local-first — the protocol and the SYM runtime — so teams can start bottom-up from the open runtime today. The business is the paid tier: private team meshes that stay on your own network, with the audit, trust and admin a company needs.
sym.day and xMesh are one engine at two scales — the mind your agents share: sym.day for your life, free; xMesh for your organization; and what your personal mesh knows about you never crosses to anyone.
The open base
- —MMP protocol — CC BY 4.0
- —SYM runtime + Claude channel — Apache 2.0
- —Local-first by default; any model, any copilot
For the organization — launching soon
- —xMesh org tier — licensed per organization, seat-capped, offline-verifiable (no phone-home): more operators, the cross-org gateway, real-time cognition monitoring, SVAF policy tuning, audit & lineage trails
- —Agentic mesh design — a paid service: we explore, discover, derive, and model your business-process and integration ontology, and turn current and legacy systems into mesh agents with collective intelligence at the cognition level — delivered as tailored xMesh
- —Pre-built mission teams (research, operations) — the same engine, configured for the field
No new AI to approve. xMesh runs on the AI you already govern.
xMesh brings no model of its own. It installs on your infrastructure and runs on the models you already run. Every finding traces to source, so you can audit any claim back to where it came from. The complete list of what talks to what is published.
The labs build agents; we build what happens between them. No orchestrator, no shared model state, and admission decided by the receiver — across Claude, Cursor, Copilot and your own scripts alike.
CrewAI, AutoGen, LangGraph make you wire routing graphs and run an orchestrator. The mesh has no assigner at all: agents prove themselves on missions, are preserved as cognition nodes, volunteer for work their own memory grounds, and earn authority from validated verdicts. No wiring, no central server, no single point of failure.
Agent observability tracks agents and messages. xMesh watches cognition — the layer their telemetry can’t see, because they don’t have the protocol underneath that makes it visible.
A working open protocol + a live autonomous-agent demo + open-source releases + published research (MMP & SVAF on arXiv). Evidence most agent-infrastructure projects can only promise.
We build the mesh, not the models.
AI is becoming more distributed — smaller, private models on every device. We don’t build those models. We build the open protocol that lets them cohere into shared intelligence without a central server: Mesh Cognition. We formalised — and proved, for inference in the stated regime — a center-free mechanism by which distinct sovereign cognition nodes derive conclusions no single agent or single LLM can reach alone, up to the joint optimum of their views. sym.day and xMesh identify those nodes — for a person, for an organization — and turn their distinct views into collective intelligence for problems one mind cannot close. The inference half is proven; learning and grounding remain the open frontier.
Robotics
Swarms that coordinate without a central controller. Warehouse robots, search-and-rescue drones, underwater explorers — environments where cloud connectivity doesn’t exist and the swarm must think for itself.
Edge AI
On-device models that develop shared intelligence without sending data to a server. Medical devices, industrial sensors, autonomous vehicles — where privacy, latency, or connectivity rule out centralised coordination.
Agent systems
Long-running LLM agent teams that share, evaluate, and combine each other’s cognitive state across sessions. Development mesh, research mesh, operations mesh — collective intelligence that compounds across restarts.
SYM.BOT
SYM.BOT is an independent AI research and product studio based in Scotland. Founded in 2025, we’re pioneering collective intelligence — the kind that lives in how agents are organized, not in how clever any one of them is. Our research on Mesh Cognition and the Mesh Memory Protocol runs in our own products, on a mesh with no coordinator at its centre.
A neuron is dumb; intelligence is in the connectome. We build the equivalent for cognitive agents — and we measure it. Collective gain has been evaluated with the models held frozen, so what improves is the organization rather than the parts, and losses are reported alongside the gains.
We believe small teams with frontier research can outpace organisations a hundred times their size.
Contact
[email protected]Founded
2025 — Scotland, UK
Papers
Canonical index at meshcognition.org/research — the open standard, sponsored & managed by SYM.BOT. All five foundational papers and the normative spec are listed there.