OrgLM.ai / White paper
State of the Union · mid-2026
Enterprise AI now agrees on one thing: context and governance, not model quality, decide whether it works. It does not agree on what to build, or who should own it. This is a fair map of the competing answers — and the open contract none of them yet fills.
The gist
Each point of view has a credible champion and a real insight. None is complete on its own. Where each one lands tells you what it leaves unowned — and the pattern points at the same hole.
| Point of view | Form | Core insight | Leaves unowned |
|---|---|---|---|
| Context graph | Thesis · VC POV | Decision traces are the strategic asset | No contract, no governance boundary, no model-interchange |
| Memory layer | Engineering | Persistence is measurable and improvable | Per-agent, session-scoped; silent on policy |
| Agent protocols | Open contracts | Interoperability is won with open specs | Transport-layer; doesn't type facts or gate actions |
| Governed execution | Platforms | Assurance belongs at runtime, not model output | Ships as proprietary platforms, not open contracts |
| Platform | Products | Demand is real; land-and-expand works now | Context siloed inside one vendor; not portable |
| OrgLM.ai + OMAP | Open contract | Unifies all five into one open, owned layer | — the slot the others leave empty |
Read on for the full version of each row — stated in its strongest form, with what it gets right and where it stops short — then our position and what it means for each actor.
§1 · The consensus
A single conviction runs through enterprise AI: the model is no longer the scarce asset — the organization's context is. The reasoning is rented and improving on someone else's schedule. What an organization knows, how it decides, and what it is allowed to do are its own.
Three independent signals frame it. Industry analysts have named context engineering a top trend, displacing prompt engineering, and project that a majority of AI projects will be abandoned through 2026 — attributing the failures to data readiness, not model quality. Widely-cited research finds the overwhelming majority of enterprise AI pilots generate no measurable return, with the same failure mode across industries: the model performs, but the context infrastructure to make it reliable does not exist. And a majority of the Fortune 100 are projected to appoint a head of AI governance within the year.
Read together, these say the bottleneck has moved. It is no longer the intelligence of the model; it is the organization's ability to feed that intelligence reliable, governed, current context — and to control what it does with it. Every point of view here accepts that premise. They diverge on the response.
The unit of investment is no longer the prompt or the model. It is the context and governance infrastructure between the organization and whatever model it rents. The open question is its shape and its ownership.
Figures here are paraphrased from widely-cited 2025–2026 analyst and research sources and should be footnoted to primary sources before external publication.
§2 · Five points of view
Stated in the strongest form its proponents would use — with its champion, what it gets right, and where it stops short. We argue with none of them here; that comes in §3.
Champion — Foundation Capital (Ashu Garg, Jaya Gupta); amplified by Box, HubSpot, Glean
The position
Enterprise software records outcomes — the final price, the approved discount — but never the reasoning behind them. The exceptions that applied, the precedent that mattered, who approved what and why live in Slack threads and people's heads, and have never been treated as data. The next generation captures them. Over time they accumulate into a context graph: a living, queryable map of how an organization actually decides. Because traces can't be reconstructed after the fact, startups capturing this layer have an opening incumbents can't easily close.
What it gets right
The sharpest articulation of why context is strategic and why it compounds. The outcome-vs-reasoning distinction is real and under-appreciated; "must be in the workflow" is a genuine moat argument; and it correctly frames ownership as the crux.
Where it stops short
A thesis, not an artifact. "Context graph" names the prize without specifying the contract, the governance boundary, or the model-interchange mechanism. The sharpest skeptics note it rebrands long-established process-knowledge management — and that software can capture traces going forward but can't reconstitute decades of tacit judgment.
Champion — Mem0 and the agent-memory tooling ecosystem; a growing benchmark literature
The position
Persistent memory is now a first-class architectural component with its own benchmarks and measurable gaps between approaches. The job is engineering: token-efficient extraction, multi-signal retrieval, temporal reasoning, cross-session identity. Done well, an agent stops starting from zero and improves across sessions.
What it gets right
Rigor. This camp turned "memory" from a hand-wave into something measured and compared, and named the genuinely open problems — privacy, consent, identity resolution, staleness — honestly.
Where it stops short
By its own admission, today's memory layer is predominantly per-agent and session-scoped: a vector store plus conversation history. That solves an agent forgetting within its thread, not the organizational, cross-agent, governed memory the enterprise needs — and it is largely silent on what the agent is permitted to do with what it remembers.
Champion — Anthropic (MCP), Google (A2A), Linux Foundation (ACP); the broader open-protocol community
The position
Fragmented, custom integrations are the enemy. Open, vendor-neutral protocols let models, tools, and agents interoperate without lock-in. MCP standardizes how applications deliver tools and context to models; A2A standardizes agent-to-agent collaboration. The win is a unified, plug-and-play substrate across vendor boundaries — a USB-C for AI.
What it gets right
Everything about the value of open standards. MCP's adoption across OpenAI, Google, and Microsoft, with thousands of community implementations, proves an open protocol can become the default fast. Interoperability is won with specs, not products.
Where it stops short
These protocols operate at the transport and tool-calling layer — how an agent reaches a tool. They are deliberately agnostic about what crosses the boundary and whether it is allowed. MCP does not type a fact, attach provenance and confidence, or gate an action against versioned policy. Not a flaw — scope. But it leaves the governance contract unspecified, sitting above them.
Champion — Oracle, IBM (Sovereign Core), a policy-aware-AI research line; Starburst on grounding
The position
Safety must move from model outputs to runtime. State — memory, summaries, pending approvals, retrieval indexes — is itself a governance surface. Long-term memory should record provenance, confidence, and classification; actions should pass hard policy gates that turn human-readable rules into auditable machine decisions. The adjacent "grounding" argument adds that retrieved context should be structured and executable — facts carrying type and confidence — not passages of text.
What it gets right
This camp independently arrived at much of OMAP's substance: typed facts with provenance and confidence, hard policy gates, attributable decisions. It is the most architecturally mature of the five, and its instinct — that the boundary, not the model, is where assurance lives — is exactly right.
Where it stops short
It ships as platforms, not open contracts. Oracle's governed execution is Oracle's; IBM's embeds policy at IBM's runtime; governance vendors sell suites. Each reintroduces a form of the lock-in the sovereignty premise was meant to escape, and none is a vendor-neutral spec another system can conform to.
Champion — Sierra (Agent Data Platform), Glean (Enterprise Graph + memory)
The position
Give the enterprise one platform that unifies everything it knows — structured and unstructured, across sessions and systems — into an intelligent layer agents reason and act on, improving with every interaction. Memory and continuity for agents, protected in a single-tenant cloud with permission fidelity and action validation.
What it gets right
Execution and proof. These are shipping, revenue-generating systems with real traction — consecutive nine-figure ARR quarters, multi-billion valuations, trillions of indexing tokens. They demonstrate the memory-plus-action layer has genuine demand now, and that land-and-expand works.
Where it stops short
The context lives inside one vendor's platform — precisely the dependency the sovereignty premise warns against. An enterprise running a dozen agents across vendors ends up with a dozen context silos, none portable. The platform thesis delivers value while quietly contradicting the consensus it is built on.
§3 · The OrgLM.ai position
Our claim is not that the five are wrong. It is that they are five views of one missing layer — and that the field's real gap is the open contract that would let all five interoperate.
The context-graph thesis describes the asset. The memory camp builds part of its mechanism. The governed-execution camp supplies its controls. The platform camp proves its demand. The protocol camp supplies the openness model. Assembled, these describe one thing: a substrate the organization owns, where typed memory and explicit policy sit beneath models that are interchangeable compute. We call it the organizational model — "model" as in a data model, a structured representation of the organization itself, which a frontier model queries but does not replace.
Shaded layers are what the organization owns. Only model access is treated as a commodity.
Each camp's shortfall points at the same hole. The context graph has no contract. Memory tooling has no policy. Agent protocols stop at transport. Governed-execution platforms are closed. The platform players silo. The connective tissue that resolves all five is a vendor-neutral protocol that does three things at once, which no existing standard does together:
Every fact crossing into a model carries type, provenance, and confidence — an open envelope, not a platform feature.
Every action is a structured request an explicit, versioned policy gate accepts or rejects — an open contract, not a proprietary runtime.
An identical request goes to any conforming endpoint — the protocol camp's openness, raised from transport to the governance boundary.
This is OMAP, the Open Memory & Action Protocol. It is defined at the level of messages at boundaries, and is deliberately silent about how a conforming system stores, ranks, learns, or reasons. The substrate behind it is a black box by design — which is exactly what lets it be open without any vendor giving away its moat.
Above MCP/A2A, not against them. The agent protocols are the transport; OMAP is the governed-fact-and-action contract that rides on top. An OMAP fact can be delivered over MCP; an OMAP action invoked through it.
Beside grounding and governed execution, made open. OMAP turns the typed-fact and policy-gate instincts those camps proved out into a spec any system can conform to, rather than a feature any one vendor sells.
We are not claiming to be first. Typed facts have a cousin in grounding; governed execution has cousins at Oracle and IBM; open protocols have a clear leader in MCP; the sovereignty thesis was named by Foundation Capital. We claim something narrower and more useful: that no one has unified these into a single open contract, and that the field needs that contract more than a sixth platform. A manifesto that ignored its neighbors would be naive. This one is built on their work and proposes the missing piece between them.
Five camps are converging, from five directions, on a layer that still has no open standard at its center.
§4 · Implications
If the organizational-model framing is right, an open contract is positive-sum for almost everyone in the field.
The field agrees that context and governance, not model quality, now decide whether enterprise AI works. It has not agreed on what to build or who should own it — and the five answers on offer each hold one piece. The context graph names the asset; memory tooling measures the mechanism; agent protocols open the transport; governed execution supplies the controls; platforms prove the demand. What is missing is the open contract that lets them compose: typed facts, governed actions, and interchangeable models, owned by the organization rather than any vendor.
That contract is OMAP, and the layer it serves is the organizational model. Building that standard, in the open, is the work worth doing now.