A question that wouldn't go away
Most projects begin with an idea. Ours began with an irritation.
I spent twenty years working on algorithmic decision systems and guiding companies through digitalisation projects. I knew the state of the art — and still, one observation wouldn't let go of me: the systems we work with every day forget everything. Every conversation starts at zero. Every insight evaporates the moment the window closes. We had language models that had read entire libraries — and no memory that reached from yesterday to today.
The question that grew out of that was both simple and uncomfortable: what would it take for an artificial system to truly remember — not for the span of a chat, but across days, weeks, years?
The personal root of that question I've told elsewhere — why I wanted to build a brain in the first place. This is about what came after: the development story itself. How, over roughly two years, that one question became an open-source memory core, an operating system for knowledge work, and three distinct research lines. I'll tell it as honestly as I can — detours included.
The long run-up
In hindsight I was doubly prepared for that question without knowing it.
One half was technical: since 2009 I'd worked on statistical models, deterministic decision strategies, and real-time processing — systems that have to decide under uncertainty in milliseconds. The other half was organisational: years in digitalisation and transformation projects, from scientific institutions up to the corporate level. Process atlases, AI strategies, decision systems across very different industries. I'd watched up close where good intentions fail inside organisations.
And there was something that, at first, had nothing to do with software: years of volunteer work in civil protection and disaster response. Incident coordination at major emergencies, technical rescue, breathing-apparatus operations. That operational reality — decisions under pressure, with incomplete information, where mistakes cost people — later shaped an entire research line. But I didn't know that yet.
This double preparation became visible later in the architecture. Building a memory system is half a mathematical and half an organisational problem. I happened to have worked on both halves.
From sketch to architecture (2024–2025)
In mid-2024 the idea took shape. If biological brains have solved the problem of remembering, why not learn from them? The first sketches were rough: a multi-layer memory structure, fed by the neuroscience I was teaching myself — Ebbinghaus's forgetting curve, spaced-repetition methods, Hebbian learning, consolidation during sleep.
Through 2024 and early 2025 those became principles: seven memory layers, sleep consolidation as a methodological cornerstone, Bayesian confidence propagation. And in parallel, something happened that I didn't at first see as part of the same project: out of my emergency-response experience, a second question formed — about rights-preserving protection mechanisms for public spaces. Two strands that would later turn out to belong together.
In 2025 the sketches became a plan. I laid out three research lines — cognitive architectures, protection of public spaces, applied AI for knowledge work — and began documenting the research as a source in its own right. I time-stamped key documents in advance via OpenTimestamps, long before any of it was public. Not out of distrust, but out of scientific discipline: whoever claims priority has to be able to prove it. By late 2025 the first memory layers were running productively.
January 2026: the build begins
Then came the phase in which architecture turned into code — and it had a force that surprised even me.
In January 2026 the productive implementation began in the current development repository. Eight architecture phases, each building on the last, in the first two weeks. The original system was still soberly called "Personal AI Brain." At the end of January it got its name: ZenAI. A private experiment became a product with a direction.
I should admit something uncomfortable here: I am a single person. That a one-person lab can work at this density is precisely because the tools we research themselves become force multipliers. One human who sets the architecture and the judgement — plus a machine that takes over the typing, testing, and looking-up. That is no magic and no trick. It's the most honest demonstration of the thesis the whole project hangs on: that good AI doesn't replace people, it multiplies them.
The month it all came together
When I look at the commit history, one month stands out: March 2026. By far the densest stretch of the whole project.
In a few weeks, what is today the core of ZenAI took shape: the cognitive architecture across many phases, the seven-layer memory, a knowledge graph, end-to-end tool use, a command-line version, a full test suite. Today more than 11,500 tests stand behind the system — an unusual number for a solo project, one I write about more closely elsewhere.
More important than the numbers was an insight that became final that March: ZenAI is two things at once. It is a productive operating system for knowledge work — and it is the test vehicle for the memory research. What the research papers describe has been running here in real use since early 2026. Research and product are not two departments but the same system seen from two angles. That remains our most important structural principle.
Going public: open source and the preprint
At the end of March 2026 I made a decision I'd thought about for a long time: to open up the memory core.
ZenBrain — the neuroscience-grounded memory architecture — was extracted from ZenAI as a package family of its own and published on npm under the Apache 2.0 licence. Nine fundamental algorithms, each backed by a peer-reviewed source; fifteen if you count the six components of the Predictive Memory Architecture that govern the memory dynamics. Why give away your hardest part? Because a memory core that people are meant to entrust with their knowledge work has to be verifiable. Trust comes not from assertion but from traceability — and the opposite of traceability is a black box.
In April came the academic preparation: LaTeX infrastructure, ORCID, replication material. In May 2026 the ZenBrain paper went online as an open-access preprint on arXiv (2604.23878), with a permanent DOI on Zenodo — fully documented, with material to reproduce the results. Follow-up publications for peer-reviewed venues are in preparation; explorations with universities I keep strictly in the conditional until written commitments are in hand. To me that's part of the integrity of an independent, self-funded lab: claim nothing that isn't proven or not yet agreed.
The stress test
Between "it runs on my machine" and "a company entrusts it with its data" lies a gap that many AI projects underestimate. In April I deliberately walked into it.
Before ZenAI was allowed anywhere near real organisations, I ran a connected series of hardening sprints: field encryption, tenant-isolated access control, defences against whole classes of attack, guardrails against prompt injection, a consent centre with disclosure and deletion, labelling of AI-generated output under the EU AI Act, a penetration-test pass. Compliance wasn't a coat of paint applied afterwards; it was a gate the product had to pass through first.
That is the least glamorous phase of the whole story — and perhaps the most important. Data sovereignty, GDPR, and the EU AI Act are not a box-ticking exercise for us; they are architecture. A system you can run entirely on your own infrastructure, from which no data flows out, is not a feature. It's a stance.
Not one product, but a lab
Somewhere in the spring it became clear that the one project had long since become more. Not one product, but a lab with three lines.
Track A — cognitive architectures. The foundation. ZenBrain is the publicly visible core, but not the only one: there are further research cores I'm working on and will publish when they're ready. About those I'll speak only when there's something solid to show — conditional until then.
Track B — protection of public spaces. The second line, CrowdGuard, is the one that grew out of my emergency-response experience. It's about rights-preserving, explicitly non-biometric protection mechanisms for public spaces — an architecture that treats civil liberties not as an obstacle but as a design principle, in line with the EU AI Act. The validation draws solely on publicly documented material; out of respect for those affected, we name no specific cases. More on the public-spaces page.
Track C — applied AI for knowledge work. This is where what people use every day lives: ZenAI as an operating system, ZenBI for business analysis in natural language, ZenSales for GDPR-compliant contact management — and the consulting with which we accompany companies along this path.
The three lines don't sit side by side by accident. They share the same substructure: the conviction that AI must be traceable, data-sovereign, and aligned with people. The research cores carry the applications — and the applications ground the research in real problems.
Today — and what remains
The most recent milestone is from this week: ZenEntry, a tool with which companies can run through their own AI roadmap in minutes. It closes a loop. The story began with the question of how a system remembers — and today it leads to helping others bring AI into their operations the right way.
When I look at the whole stretch — from the algorithmic beginnings through the 2024 question to what runs productively today — the throughline isn't a technology. It's a stance: claim nothing you can't prove; build the foundation correctly before building on top of it; multiply people rather than replace them.
This story isn't finished, and it isn't fully told yet either. In the coming posts I'll pick out individual chapters — the stress test, how we deal with AI, the separate research lines — and tell them in more detail. Step by step. The sober, fact-checked version with all the dates is always on the research page; here I tell the journey as it felt.