Starling Memory Works has made its Universal Cognitive Architecture available as a free standard, proposing an enterprise-AI model in which organisational knowledge remains in a governed repository rather than depending on the memory of a particular language-model provider.
The Princeton-based company calls the resulting architecture a Domain Language Model, or DLM. Its definition combines a stateless language model with a sovereign repository containing the organisation’s approved knowledge.
The model is treated as a replaceable reasoning engine, while the organisation retains control of the information on which that reasoning depends.
Starling’s Universal Cognitive Architecture, or UCA, provides a classification system intended to give organisational knowledge permanent semantic addresses. Rather than asking an AI system to search a large collection of documents each time it needs information, the architecture is designed to direct the model towards a defined location in the repository.
The company has made the standard available under Creative Commons, allowing organisations and technology providers to implement the approach independently of Starling’s commercial platform.
Starling uses Model Context Protocol to connect AI systems with the repository. MCP is an open standard for linking AI applications with external data, tools, and workflows, giving developers a common way to expose information and functions to models.
Starling founder and chief executive Chris Kincade argues that the ownership and structure of organisational memory should be treated separately from whichever model is being used to reason over it.
“Sovereignty is not a feature or add-on. It’s how your knowledge is stored — and it’s the decision you have to get right before everything else,” Kincade said.
The company divides information into two layers. Its Org Library is intended to hold authoritative organisational memory, including decisions, standards, policies, and established ways of working, in versioned plain-text records.
A separate resources layer contains incoming references and rendered outputs such as PDFs, presentations, and other formatted documents. Starling’s argument is that the underlying organisational knowledge should remain independent from the document in which it happens to be presented.
The approach addresses a growing enterprise-AI problem. Businesses increasingly use several foundation models, productivity assistants, and specialist applications while important information remains spread across document repositories, collaboration platforms, email, databases, and individual accounts.
When useful context becomes embedded in one provider’s proprietary memory or workflow, changing models can force organisations to reconstruct part of the knowledge environment on which the system depended.
Starling proposes the opposite structure: keep organisational memory under the company’s control and allow different models to reason from the same governed source.
That separation has become more relevant as companies examine concentration risk around AI providers. Availability, pricing, policy, product changes, or technical disruption at one model provider can affect workflows built around that platform even when the organisation’s own information remains unchanged.
Government investment is addressing another part of the sovereignty question. The UK’s £1.1bn AI Hardware Plan includes £750m for a national AI supercomputer and wider measures intended to strengthen domestic computing, chip capability, and technical skills.
Starling is addressing sovereignty at the organisational-knowledge layer rather than the compute layer, concentrating on ownership, classification, permissions, and portability.
The architecture still has to compete with established approaches including retrieval-augmented generation, enterprise search, vector databases, knowledge graphs, model fine-tuning, and proprietary memory functions. Those technologies address different parts of the same problem, and many organisations are likely to use several approaches together.
UCA’s distinguishing proposition is that humans should assign durable meaning and location to important organisational knowledge rather than delegating its structure entirely to search algorithms or a software vendor.
That puts governance at the centre of implementation. A repository provides a reliable source of truth only when organisations determine who can create, change, approve, and retire information. Poorly governed knowledge remains unreliable input even when its technical address is permanent.
Starling’s commercial platform sits around the open standard, creating a clear distinction between the architecture it wants others to adopt and the product it sells to organisations that do not want to assemble the system themselves.
Making UCA freely available means companies can test the underlying approach without buying Starling MX. Starling’s commercial case will therefore depend on whether organisations see value in a managed platform for implementing and governing that architecture rather than building comparable systems through existing data and AI infrastructure.
As enterprise AI becomes increasingly model-agnostic, control of context, permissions, provenance, and organisational memory is emerging as an infrastructure decision in its own right. UCA is Starling’s attempt to turn that layer into an open architecture rather than a vendor-specific feature.
You must be logged in to post a comment.