Starling open-sources AI memory standard for organisations

Starling Memory Works has published its Universal Cognitive Architecture as a free open standard, defining what it calls a new class of AI system: a Domain Language Model.

The Princeton-based company also launched Starling MX in beta as the first platform built on the architecture. Pricing starts at USD $99 a month for the first seat, with additional usage charged separately.

Universal Cognitive Architecture, or UCA, is designed to let organisations store knowledge at fixed semantic addresses that different AI models can read. Starling says this lets a model retrieve information by coordinate rather than by search, while keeping the underlying repository under the organisation's control.

The company presents that approach as part of a wider debate over AI sovereignty. Much of that discussion has focused on where models are hosted, where data resides and who owns computing infrastructure. Starling argues that governance of organisational knowledge has also become an operational issue.

Chris Kincade, Chief Executive Officer and Founder of Starling Memory Works, said the current debate misses a more basic question about how knowledge is stored and governed.

"The current conversation on AI sovereignty focuses on infrastructure - where models run, where data sits, and who owns the compute. It misses the real story. Sovereignty is no longer a trade-off. Governed memory produces better intelligence than what AI generates. It detects drift and keeps every conversation current - giving businesses the very real gift of context," Kincade said.

Starling defines a Domain Language Model as a stateless large language model that reasons over a separate repository, rather than storing an organisation's working memory inside the model itself. In practice, this means businesses can switch between models for different tasks while keeping the same source of organisational knowledge.

The architecture uses Model Context Protocol, the interoperability standard developed by Anthropic, to connect a model to the repository. Starling says UCA directs the model to the relevant address in that repository, removing the need to search a broad information store each time a query is made.

Memory structure

A central part of the system is what Starling calls a dual repository. One side, the Org Library, is intended to hold versioned plain-text records of decisions, standards and processes as the main source of truth. The other, called Resources, holds reference materials brought into the system and documents produced from it, such as PDFs, slide decks and other formatted files.

The distinction reflects a technical and commercial judgment about how AI systems handle information. According to Starling, plain text is cheaper for models to read and reason over than formatted documents are to convert and process. The repository is therefore designed to keep working knowledge in a text-based layer and treat documents as outputs rather than the primary record.

Starling says the structure is intended to reduce dependence on software vendors' proprietary storage methods. By publishing UCA under Creative Commons, it says organisations and software providers can build on the standard without licensing or approval from Starling, while keeping a knowledge base in plain text rather than in a closed system.

Kincade linked that position to a broader view of software design and human control over business systems.

"Where generations of enterprise software trained humans to speak machine, Starling trains machines to speak human - building organizational intelligence that lives on your infrastructure, governed by your people. 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.

Early users

Starling says organisations in marketing, personal coaching, construction management and legal services have already used the Starling System Library to build their own Domain Language Models. Examples include SelfOS, a personal operating system built by Brady Patterson, and QuillOS, a copy production system used by marketing agency Wrks Online for client work.

Starling also says it used the architecture internally while building its own platform. It ran the build through a single governed semantic address known as SY50 Platform Master, which it presents as evidence that the classification system can scale across a complex project.

The company was founded by Kincade, who developed UCA and what Starling calls its Cognitive Design methodology. The engineering team includes Chief Technology Officer Jeremy Decker, whose background spans financial services, healthcare and telecommunications; Vice President of Engineering Zuhayr Tariq; UI and Interface Lead Ahmed Khalid; Infrastructure and Integrations Lead Andrey Gubanov; and Repository Lead Harol Garcia.

Starling says organisations can begin using the system through a conversational process rather than software development work. For larger businesses with established materials, a core layer of what it calls Constitutional Memory can be created in a few hours. Earlier-stage companies can use a built-in process to define values, business model and brand strategy before the system proposes a wider knowledge structure.

The publication of UCA adds another entrant to the growing market for tools designed to separate enterprise knowledge from any single AI model vendor. The issue has gained prominence as governments and companies scrutinise model access, control and resilience, including in markets such as the UK, where policymakers have announced a GBP £1.1 billion AI sovereignty plan.

Starling's argument is that the repository, rather than the model, should be treated as the long-term asset. "The model forgets; the organization remembers."