Ontology. World model. Company brain. LLM wiki. Four names for one thing: your company's data, centralized and contextualized so machines can work with it. The interesting question is not what to call it. It is why this layer might be the only part of your AI stack that lasts.
Rent the intelligence. Own the memory.
Start with an inversion. The conventional bet says value lives in the AI layer—the best model and the best prompts. The counter is simpler: models are swappable and chat apps are rebuildable. What cannot be swapped is the layer underneath—your accounts, history, decisions, and unwritten rules.
The slogan is that the AI layer is a commodity and the ontology is the moat. One honest footnote: full commoditization is often asserted, not demonstrated. What still survives is softer and more useful: context outlasts models.
What this layer actually is
It is the nouns and verbs of a business written as structure. Customers, deals, projects, meetings, decisions—the entities. And the relationships: this meeting concerned that account, and that decision changed it.
Most companies still store this in documents. A fact trapped in one file serves one workflow. Attached to an entity, it can serve every workflow that touches it. Prose is good for judgment. As operational state, it is a dead end.
Connectors are not context
“But my tools are already connected.” The tell is familiar: teams with live transcription connectors still paste the transcript in by hand. A connector moves data, not context. It fetches the file, not what the file means.
Resolving who, what, which account, and whether the fact is still true is its own stage. Skip that stage and the model re-derives your business from raw exhaust on every task.
Drift is the silent killer
Even if you hand-build that context once, it rots. The pattern: the system works beautifully for six months. Then a person leaves, a goal shifts, or a process is quietly replaced. The context stops matching the company. Nothing throws an error. The tool becomes subtly wrong—and then quietly abandoned.
The problem was never capability. It is maintenance.
Architecture is a loop, not a warehouse
Events are captured where they happen: an email lands, a call ends, a task closes. Each is contextualized on arrival—who is involved, which account, whether anything changed—then stored as structure that can be retrieved later.
Run that loop long enough and experience accrues. Outcomes and corrections become events too. That is the twenty-year operator: compounded trial and error, made queryable. Ten agents with ten private hacks make every new agent more expensive. A shared substrate inverts the curve—the tenth agent gets cheaper.
Abstention belongs in the context layer
A real system knows when not to act. Deterministic policy should live as context: which actions are reversible, which touch money, and where authority ends. High risk or low confidence should stop the agent and escalate to a human.
Put abstention in the context layer so swapping the model never silently changes what the system is allowed to do.
Where the pitch thins
First, vendor-capture fear often blurs transit with retention. Keep the hygiene; skip the panic. Second, “self-annealing” is a label, not a mechanism—who validates the update? A map that corrects itself wrongly is worse than a stale one because it is trusted. Third, almost nobody prices ontology maintenance: schema work, entity resolution, and keeping the edges true. The bet that automation absorbs all of that is still unproven.
The honest hybrid
Two substrates, deliberately:
• Postgres holds what agents do—entities, events, relationships, and state—with provenance and validity windows so staleness is queryable.
• Obsidian holds what agents understand—decisions, rationale, and nuance.
The loop closes under one rule: automation proposes; validation promotes. Nothing becomes trusted knowledge until a human, or accumulated evidence, says so.
The moat, stated honestly, is not magic. It is your data plus how it relates, kept in a layer you own, maintained by a loop you control—with the humility to stop when it is not sure. Models will keep changing. That is fine.