Chris Kincade spent twenty-five years building across media and tech: first packaging books for major publishers like McGraw-Hill, then designing platforms for social bookmarking and news personalisation, ultimately shaping the frameworks that defined Cognitive Design.
Starling Memory Works is where those skills converge. The company started with what were, effectively, operating manuals for AI systems, graduated to an AI-agnostic knowledge repository and ultimately became an “Org Brain” constructed from those twenty-five years of business blueprints. The platform gives organisations a permanent, governed memory that any AI can work from – so the AI finally knows the business, and the business owns what it knows.
Kincade first debuted the Hierarchy of Organisational Values at a January 2023 summit co-hosted with Ari Meisel, spending the next two years refining it with clients before adapting it for AI. He wrote Starling’s operating system himself, in English rather than code, before bringing in an engineering team to build it into a platform. That team included CTO Jeremy Decker, who previously designed a back-end for Mastercard, divisional leaders Zuhayr Tariq, Ahmed Khalid and Andrey Gubanov, and Harol Garcia, who joined in June to build the company’s own repository.
He shared his thoughts with us on why he thinks most AI tools have the ownership problem backwards.
What problem were you trying to solve, and why a new approach rather than an improvement on existing tools?
Every AI tool on the market shares one assumption: the intelligence lives on the machine’s side, accumulating inside a platform you don’t own.
The fix is organisational, not technical. We have to decouple the AI system from our memory, which includes daily work (i.e., active memory). Whatever the model needs to provide contextualised assistance has to live in a governed, plain-text repository the organisation owns. Any ‘permissioned’ AI can read and write to it, but the repository has to be so well-wired that retention is not even necessary.
That’s why it had to be a new approach. The industry connects AI to knowledge through derived copies – embeddings, vector databases, retrieval pipelines – shadows of the truth that start drifting the moment the truth changes. We eliminated the category. In Starling, the library itself is the cache. Context is assembled fresh from your canonical source every session. Read, reason and forget – the next session reads the same source again. Nothing goes stale, because there is no copy to drift against. There is only the source of truth.
You’ve described most AI leaders as having an Oppenheimer complex. What do you mean, and how does Starling differ?
This summer, a thousand AI workers, including prominent influencers, petitioned Washington to regulate their own field. When the people building a technology ask the government to restrain them, they’re admitting they haven’t built restraint into the technology itself.
I’d rather keep the machines on our side through architecture and design – which means, in practice, that people hold every important switch. Nothing the AI writes becomes company truth until a person approves it. Every answer shows where it came from and who verified it. A written code of conduct governs what the AI may and may not do.
The AI can do the drafting and the finding; only people do the deciding. That’s what putting humans at the centre actually entails. It’s not a policy statement, but switches only people can flip.
Why make Universal Cognitive Architecture free, and what do you hope it enables beyond Starling?
Governance shouldn’t be proprietary, nor should the ability to own your memory. The standards that changed the world were open: the Dewey Decimal System didn’t license libraries; HTML didn’t paywall the web; and Git didn’t meter commits. UCA (Universal Cognitive Architecture) gives every piece of organisational knowledge a permanent, human-readable address, and that only becomes infrastructure if anyone can build on it without asking me for permission.
UCA is a free standard with instructions on how to wire it into a Claude project. I hope it enables a new profession – Cognitive Designers, able to structure organisational memory for human governance. The first builders shipped their own systems on the published framework before we’d launched anything commercial – that’s part of the point. If UCA only ever powered Starling, we’d have a killer app. Our ambition is a retrieval and assembly standard people adopt because it’s fast and effective, not just sovereign.
In short, we want to launch a movement. Honestly, if you follow the logic all the way through, we’re pursuing an alternative to AI having instant access to all human knowledge. That’s what the massive data centres are needed for. The DLM solves the big compute question on an organisational basis. With it, companies find their knowledge the same way whether they have 28 documents, or 28,000 – by address, not search.
Data centres are needed to synthesise across knowledge bases built on a SaaS workflow, with scores of document, image and database types, hundreds of APIs and thousands of applications. UCA introduces an AI-native workflow. Knowledge lives in lossless markdown. What leaves the building gets converted when you need it.
You describe knowledge stored inside AI platforms as an operational risk. What does that mean in practice?
Well, first and foremost, when AI trains on your ‘special sauce’, your unique values and protocols become part of the new baseline. Yes, you can write ‘no training and storage’ into a Service Level Agreement, but without memory, the AI is a brilliant mess.
Two events this summer showcased the risks of depending on AI platforms. In June, a government directive pulled frontier models from global access overnight. In the 18 days it took to restore access, any working knowledge inside the platform was unreachable. In a separate incident, a security team in the middle of a crisis asked a commercial model for help and was refused by the model’s own guardrails.
The first is a memory risk: knowledge you don’t hold can be held from you. The second is a cognition risk: rented intelligence can decline you. A Domain Language Model survives both. We define a DLM as stateless, session-based, and redundant AI systems reading and writing to a sovereign repository. If any one AI system is unavailable, the system can switch to another.
