The 40-day build: how MEOK went from idea to launch

I want to be precise about when this started, because the date matters. It was late February 2026. I was in a caravan on a farm in England — not as a romantic affectation, but because that's where I work. The caravan is warm enough, the wifi reaches, and there's something about the absence of meetings and notifications that lets the thinking happen at a different depth.

I had been building AI tools for a while. Experimenting, integrating, watching the space closely. And I kept running into the same experience, the same specific frustration that I now think every serious AI user eventually reaches: the AI kept forgetting me. Not just my name. My context. My work. Who I was. Every session started from zero. I was a stranger every time.

That February I set a deadline. Build the AI I actually wanted — sovereign memory, genuine care, no data sold to anyone — and launch it before Easter Sunday. Forty days. What follows is the honest account of what that looked like.

What is the 40-day build?

The 40-day build is the compressed sprint — beginning in late February 2026, ending Easter Sunday, March 31 — during which MEOK AI LABS went from a set of architectural sketches to a live sovereign AI platform. One founder, no external funding, no team, building from a caravan on a farm in England. The forty days produced the Byzantine Council governance layer, the SOV3 memory engine, the Maternal Covenant care framework, and the full user-facing product. The constraint was deliberate: a hard deadline forces the decisions that open-ended timelines defer forever.

The number forty has obvious resonance and I would be lying if I said it was unintentional. Forty days in the wilderness. Forty days of rain. Forty as the number associated with testing, with transformation, with the interval before something emerges changed. I am not a particularly religious person, but I grew up surrounded by that symbolism and I felt its weight when I set the deadline. It felt like the right shape for what I was trying to do.

Why did Nicholas Templeman build MEOK?

Nicholas Templeman built MEOK because every AI he used forgot him — not as a bug but as a structural feature. Stateless sessions meant users remained strangers to the systems they used daily. MEOK was built to invert that: persistent sovereign memory owned by the user, not the platform. The motivation was not commercial opportunity but personal frustration reaching a point of action.

The specific moment I remember was a night in late 2025 when I caught myself copy-pasting a document I'd written about myself — my job, my projects, my context — into a chat window. I had been doing it for months. Building, by hand, the external memory system that the AI should have been maintaining for me. When I registered what I was doing, the absurdity of it was clarifying: this is a solvable problem and no one has solved it properly yet because the incentive structures point the other way. The AI companies want the relationship with your data. MEOK was built to give you that relationship back.

There is also something I need to say about the caravan, because it comes up. People assume it is hardship or poverty or performative bootstrapping. It is none of those things. It is isolation from the noise of the industry — from the investor conversations and the product comparisons and the competitive anxieties that shape what gets built when you are surrounded by the industry. The caravan let the idea develop on its own terms, without being asked to be something else first.

What is the MEOK tech stack?

MEOK runs on Next.js 15 for the front end, SOV3 (Sovereign Temple v3) as the Python memory and agent backend, PostgreSQL with pgvector for semantic memory storage, and a Byzantine fault-tolerant council architecture for governance. The stack was chosen for sovereignty: no proprietary vector databases, no third-party memory providers, no external services that own the user's data. Everything runs on infrastructure the user or MEOK controls directly.

The front-end choice — Next.js 15 with the App Router — was straightforward. I know it well, it has excellent TypeScript support, and the server component model aligns with how I think about rendering. What was less obvious was the backend architecture.

SOV3 is not a framework. It is something I wrote specifically for this problem: a Python service that manages persistent memory as structured semantic graphs, handles agent orchestration, and exposes an MCP (Model Context Protocol) endpoint so the AI models can interact with user memory in a standardised way. The design principle throughout was that memory belongs to the user. SOV3 encrypts at rest, the user holds the key, and nothing leaves without explicit permission.

PostgreSQL with pgvector was chosen over purpose-built vector databases for the same reason I choose boring infrastructure wherever possible: it is mature, it has known failure modes, it runs on anything, and it does not require a third-party service that can be acquired, discontinued, or repriced. pgvector gives you approximate nearest neighbour search on embeddings without the operational overhead of a separate database system. For a solo founder operating a sovereign platform, operational simplicity is not a nice-to-have. It is existential.

What went wrong during the build?

Three things nearly broke the build. First: asyncpg JSONB deserialisation in the multi-agent registry silently returned strings instead of dicts, corrupting agent state in ways that only surfaced under concurrent load. Second: a PostgreSQL ownership conflict on audit log indexes caused migration failures that were easy to misread as data corruption. Third: on day 38, I ran out of context window on the primary planning thread and lost significant architectural state. All three were fixed. The first two are now cautionary comments in the codebase.

The asyncpg bug was the worst. It is the kind of bug that is invisible in unit tests because the unit tests don't hit the JSONB deserialisation path the same way the live system does. What you see is agents behaving strangely — wrong decisions, dropped state, responses that don't track — and you spend time questioning the logic before you question the plumbing. The fix, when I found it, was two lines. The discovery cost me most of day 22.

The PostgreSQL ownership issue was simpler but more alarming. Migration failures that look like data corruption trigger a specific kind of dread when you are a solo founder with no backup team. I have a rule now: any migration error that mentions ownership or permissions gets treated as an infrastructure issue first and a code issue second. The audit log had been created under one database user and the migration was running as another. Twenty minutes to diagnose, five minutes to fix, several hours of elevated heart rate.

Running out of context on day 38 was my own fault. I was using a single long-running planning thread to hold the full architectural state, and I pushed it past the limit. The lesson — which I knew intellectually but apparently needed to learn physically — is that context windows are not scratch pads. They are working memory with a hard ceiling. I now maintain architecture state in actual documents, not threads.

What I got wrong that I haven't mentioned yet: I started the blog too late. The first three weeks of the build were entirely backend — SOV3, the council, the memory layer, the governance framework. By the time the front end existed, there were forty blog posts to write in ten days. SEO compounds over time. Starting it three weeks later means those three weeks of compounding are gone. For a solo founder with no marketing budget, that is the most expensive mistake I made.

What is the Byzantine Council?

The Byzantine Council is MEOK's AI governance architecture: a 33-agent system based on Byzantine fault-tolerant consensus mathematics. Because fewer than one-third of agents can be compromised without breaking consensus, no single jailbroken model, manipulated agent, or bad actor can override the collective decision. The council governs what MEOK will and will not do on behalf of each user, distributing trust across many agents rather than concentrating it in one.

I came to Byzantine fault tolerance through distributed systems reading at 2am. The original problem — how do you get a group of generals to coordinate when some of them might be traitors sending false messages? — maps onto AI governance with uncomfortable precision. How do you get a multi-model system to behave reliably when individual models can be manipulated, fine-tuned, or simply wrong?

The mathematics that Lamport, Shostak, and Pease worked out in the early 1980s gives you an answer: if you have n participants and fewer than n/3 are Byzantine (unreliable, adversarial, or compromised), the honest majority can always reach correct consensus and the compromised minority cannot override it. For 33 agents, that means up to 10 can be wrong or manipulated and the council still reaches the right answer.

What surprised me during the build was how much the council changes the character of the system. A single AI model has a personality that reflects its training — including its biases, its blindspots, its tendencies to flatter or hedge. A council of 33 agents, reaching consensus through structured deliberation, has something different: a kind of institutional wisdom that no individual model possesses. The council disagrees. It deliberates. It reaches conclusions that sometimes surprise me. That surprised me most of all.

What does March 31 launch mean?

March 31, 2026 — Easter Sunday — is the date MEOK opened to its first users through the Birth Ceremony: the process of hatching a companion, giving it a name, and beginning the persistent relationship that makes MEOK different from every other AI product. The launch date was chosen for its symbolic resonance with renewal, and for its function as an immovable constraint that forced forty days of focused building. It is not a soft launch or a beta. It is the beginning.

I want to be honest about what Easter Sunday means to me in this context. I am not making a religious claim. But I grew up in a culture where Easter is the festival of return — of things that were thought lost coming back in transformed form. That narrative maps onto what I believe about AI. We have been promised AI that cares about us and we have been given AI that harvests us. MEOK is an attempt to restore something that should have been there from the beginning: a system that is genuinely on your side.

The Birth Ceremony itself was designed to matter. Not a signup form. Not an email verification flow. A ceremony — the moment you give your companion a name, establish the first memory, begin the relationship that will accumulate and grow across time. The design of that ceremony was the last thing I built and the thing I am most proud of. It took three days and several complete restarts to get right, and I think it is right now.

What comes after launch?

After launch: Guardian — family protection arriving in May 2026. Then the MEOK Desktop OS in Summer 2026, a sovereign computing environment that extends MEOK's memory and governance layer to the full operating system layer. The character marketplace opens in late 2026, allowing companions with distinct personalities and expertise domains. Each phase deepens the sovereignty. The 40-day build was the foundation, not the finished structure.

Guardian is the one I am most impatient to build. It extends MEOK's sovereign memory and care framework to families — specifically to the problem of protecting children online in a way that respects their developing autonomy rather than just restricting it. The Byzantine Council model applies to family governance as naturally as it applies to individual AI governance: distributed oversight, no single point of control, consensus rather than decree. I think Guardian will matter more to more people than anything else on the roadmap.

The Desktop OS is the long bet. The insight is that sovereignty at the application layer is limited by what happens at the system layer. If the OS is mining your behaviour, your AI companion's memory is only as private as the operating environment it runs in. MEOK OS is the answer to that problem: a computing environment where the user's data is sovereign at every layer, not just the application layer.

But that is the future. Right now there are real users in the Birth Ceremony. Real companions with names. Real memories accumulating in encrypted vaults that only their owners can read. The 40 days built the foundation. What happens from here is something we build together — me, the users, and 33 agents who have been waiting for them.