From twenty years of enterprise integration to Organizational Cognition.
Integration moved information. Automation moved work. The question both left unanswered is how organizations become better at making decisions. Big programmes still stall because the work hits blockers in every function it touches, and no one function can clear them alone. xMesh is built to work them together — your experts signing off, two decades of integration judgment in the room — so what the work establishes is captured with its owner, and holds when the organisation changes shape.
Organizational Cognition didn’t begin with large language models.
It began with twenty years of enterprise integration — distributed systems, semantic data models, enterprise architecture — across investment banking, government and healthcare. The technology kept evolving: ESBs to APIs, SOA to event streaming, orchestration to agentic workflows. Every generation moved information better than the last.
And every generation left the same question unanswered: how do organizations themselves become better at making decisions?
Those disciplines taught us how organizations move information. AI forced the new question: how do organizations move judgment? Everything we build today is our answer.
moved information between systems.
moved work through processes.
move work through AI agents.
coordinates judgment across all of them.
Enterprise integration taught us how to connect systems. Organizational Cognition asks how connected systems become better organizations.
xMesh is our enterprise implementation of Organizational Cognition. It applies the open Mesh Memory Protocol (MMP) so sovereign AI systems can selectively exchange cognitive influence across enterprise boundaries — built on everything two decades of integration taught us, designed for AI-native organizations. xMesh is launching; the engagement below is how it enters a real organization: observationally, criteria agreed before anything runs.
- Legacy integration
- Data integration
- Application integration
- Semantic integration
- Agent integration
- Organizational Cognition
It was never one blocker.
Moving the data was never the hard part. Meaning is — and not because it is hard to write down. It is that “store the meaning centrally” has no well-defined target: one term carries a different definition in each line of business, and each is correct where it is used. So what stalls a programme is everything around the data: blockers spanning business, engineering and risk, each owned by a different function, none of which close alone.
The common ones are named here. A real run surfaces as many as the work touches — each with its own owner, and none of them closing alone.
The common ones are named here. A real run surfaces as many as the work touches — each with its own owner, and none of them closing alone.
No one function can clear these, and clearing them one at a time is what makes programmes run for years. They have to be tackled together, by the people who own them — which is exactly the shape a mesh is built for.
Why integration is structurally a mesh-cognition problemResearch, implemented. Expertise, in the room.
Clearing blockers across every function at once needs two things no tool has had together: a way for autonomous parties to reach an understanding neither held alone, without a centre — that is our mesh-cognition research, implemented by xMesh — and the judgment to know what to look for, which a SYM.BOT operator brings from two decades of enterprise integration. Agents propose; your experts stay the authority.
Mesh cognition
Autonomous parties reach an understanding neither held alone — cognitive coupling, without a centre. Published as the Mesh Memory Protocol (MMP), an open specification.
xMesh
The research, built and running. Deployed per team, on your systems and your existing AI providers — each team keeps its own mesh and they federate, so a cross-team dependency stops being a queue.
Founder-led, in the room
The founder’s two decades of enterprise integration — investment banking, government, healthcare — brought to your engagement directly: a SYM.BOT operator runs it, deciding what to look for and what to ask next. What proves out there is what we encode into the agents.
No black box. Agents propose and cross-check; every decision is signed off by its owner. Nothing changes your systems on its own, and it is auditable throughout.
Runs on your infrastructure, your systems and your existing AI providers — no new AI model or provider to approve. MMP is an open specification.
The map assembles from one real initiative at a time, each increment usable on its own rather than only at the end of a multi-year programme.
Knowledge is captured with its owner and lineage, so reorganizations and modernization waves don’t send you back to zero.
Recovery without a cloud.
Machines are the most disposable thing in your estate; the knowledge this engagement establishes is not. Every entry in the register is signed by its owner and content-addressed, so a rebuilt laptop or a refreshed VM recovers the record from the team’s own mesh — nothing phones home, nothing sits in our infrastructure. And identity cannot be faked back: an agent that loses its key cannot impersonate its former self, and the system refuses to start a look-alike rather than let one in.
Preserved outside the machine — your vault, your HSM — and restored before the agent rejoins the mesh. It is the one thing no one can give back.
What your experts established stays provable — same author, same hash — through every reset.
Why no backup product can say this: a backup restores data by trusting the custodian that held it. Here there is no custodian — the team’s mesh already holds what it admitted, because admission is how the team works together — and every recovered block proves its own integrity and authorship. What a backup can never restore is identity: only the preserved key proves the agent is still itself. That is a property of the mesh-cognition architecture, not a feature added to it.
What you keep.
The output of the first three weeks, on one real initiative — not a product deliverable.
No named owner for the target’s customer domain
Target · unassigned → Identify and confirm the data owner
“Active customer” means different rules per line of business
Both LoBs → Agree the canonical definition
Downstream processing runs on-prem and cannot consume the cloud-held global data
Platform ✓ Provide a supported path from the cloud store to the on-prem consumer
The change needs three LoB teams to ship, and none has it on this quarter’s plan
LoB engineering leads → Commit the three teams to one sequenced plan, or agree an interim path that needs only one
The target exposes no interface for the customer records in scope
Platform · Integration → Agree an export contract
The target has no idempotency key — a retry double-posts
Target · Platform → Agree an idempotency key and retry semantics
Failed records have no dead-letter destination or replay path
Platform · Integration → Agree a dead-letter destination and replay procedure
No transaction boundary across the two systems — partial failure leaves them divergent
Both platforms → Agree the consistency boundary and compensation
status = 07 has no crosswalk to the target’s SUSPENDED
Target · Risk → Map 07 → SUSPENDED
Source addresses unvalidated — 12% fail the target’s format
Source · Data engineering → Cleanse and normalize before load
PII fields in scope require a DPIA before use
Risk · DPO → Complete DPIA; approve fields
EU customer records can’t leave region for processing
Legal → Process in-region or restrict fields
… and whatever else the work hits — a run surfaces blockers in whatever functions it touches, including ones not listed here.
Columns are detection maturity — how far xMesh’s detection of that layer has shipped. “Elicited” marks a finding established from the owners’ answers rather than observed in a run.
Illustrative — the blockers, owners and resolutions are examples, not a real customer’s data.
The engagement leaves xMesh running — a 30-day evaluation, included.
It is already deployed and approved inside your perimeter to do the work. Rather than take it out, we leave it — and the 30 days start at your first use, not the day we finish.
Your initiative team, plus one evaluator per function — ten in total.
One scheduled review inside the 30 days. Not production support.
It stops. Your data stays with you. Removal is documented.
Observational, as in the engagement — no binding changes, and no phone-home.
xMesh is launching. Agents assist; your experts stay the authority and sign off every decision. The engagement is observational — it makes no binding changes to your systems.