build a local library, hire a librarian.

My grandfather carried a small black notebook in his pocket and wrote everything in it. Notes, records, the things he did not want to lose. He worked in finance and ran ventures of his own, an import business and a small manufacturing company. He kept his own practice as a certified public accountant. He was one of the most structured people I have ever known. He documented his world by hand in pen and pencil, decades before anyone used the word data the way we use it now.

He kept a library at home that was alive. Every week the post brought him the Official Gazette, the government's record of new laws and amendments. It came as small loose pages. He taught me to file those pages into hundreds of folders of regulation and commerce, inserting the new, removing the outdated, keeping the whole thing current. I had no name for it then. I know the name now. It was version control by hand, on paper. His library was a living index of an entire country's law. He maintained it the way you maintain a system.

He taught me the ledgers too, where every figure has to reconcile. Before any of that, I kept my own records as a boy. Day after day I wrote the weather on squared paper, the temperature and the pressure. I wanted to analyse it later. By the time I was eleven or twelve my favourite subject was commerce, taught at an ordinary state school: companies, accounting, the architecture of how an organisation actually works. I did not realise how rare that was, nor that I was forming a habit that would outlast every job I ever had.

The habit kept finding new tools. For years I ran my whole working life inside mind maps, every task and document and relationship laid out like a nervous system. I kept them on my own machine because that was simply where my mind lived. Across a large company spanning three continents I met perhaps two or three other people who worked this way.

Then I spent years in cloud. Cloud taught me the lesson under everything I am building now. Your data, your privacy and your sovereignty are not details. They are the whole question. The more I helped move the world's information into someone else's data centre, the more certain I became that the most important things should stay where you can put your hand on them.

So one ordinary day I found Obsidian and plain Markdown. I read just enough about Llama and MLX to see that capable models would one day run on hardware I owned. Something settled. If I already kept my whole working memory in plain text that belonged to me, why should the processing that reads it belong to anyone else? This was more than a “second brain”. Why not a twin built from my own life, one I own.

Let me say plainly what I mean, because the term is already being pulled in several directions. An AI Twin is a persistent system that carries your knowledge, memory and intent across whatever AI models you use without scattering your context into disposable chats. Three things hold it up. It remembers across sessions, so nothing is lost when you close the window. It rides across whichever model you choose, so you are never locked to one vendor. And it lives on infrastructure you control, so you own it the way you own a notebook.

It helps to say what this is not. It is not the operational twin enterprises buy to simulate a company. It is not the physics twin engineers build of a factory floor. It is not a cloud assistant keeping your data on someone else's servers behind a residency promise.

Look closely at the arrangement almost everyone now lives inside. ChatGPT, Gemini and Claude all remember you now. It is genuinely useful. But it is their memory of you, built on their side, kept in their data centre, working only inside their own product. They decide what it keeps. They can rebuild it overnight. They can reset it on a pricing decision. The processing was always theirs. Now the memory of you is theirs too.

So why not simply let them remember you, since they already do? Because the question is not whether a memory exists. It is whose it is and how it works. Theirs is locked to one product compared to a vault of your own which works with any model. Theirs is inferred from your chats; your vault holds what you chose to write, in words you can read. Theirs changes when they change the product; your vault changes only when you do. I build my own twin with Claude. Even so its memory of me lives on Anthropic's side, not mine. The tool can be excellent and the arrangement still be backwards. Underneath it all sits the thing that actually matters, continuity. Models will come and go. The accumulated picture of who you are should outlive every one of them.

The model is not the precious part. I am not even sure models will keep getting cheaper or more available to everyone. An “Intelligence Caste” is the more likely future, intelligence rationed by who can pay. I make the case in my manifesto at imbros.ai. The precious part is the thinking behind the questions, and the memory it leaves. That should be yours.

I want to be precise, because it is easy to overclaim. When you rent a cloud model your notes still travel to it, so renting is not privacy. What you own is the original, the canonical copy on your disk in plain text. It is not locked inside a vendor's store. You can leave any provider on any day and take everything with you. The day nothing has to leave the room at all is the fully local version. That part is still ahead of me.

And it is not only for people like me. A teacher with years of lesson plans, a lawyer who has reasoned through a thousand cases, a parent holding the long thread of a family, each is already building a library worth keeping. The more of your thinking you pour into something that is yours, the less it matters whose processing you bring in to work it.

Now the honest part, because this is a lab notebook and not a launch. What exists is real. I live inside an Obsidian vault every day, around four hundred notes in a numbered system. Decisions logged, projects, daily entries, the direct descendant of my grandfather's library. I have even handed its daily upkeep to AI, because expecting everyone to learn Markdown by hand is the surest way to keep this a hobby for a few. What I have not done yet is point a model I fully own at the library I fully own and close the loop. That is the work in front of me. I am doing it in the open.

Right now agentic AI is the loudest thing in the room, software that promises to go off and act for you on its own. I think it is the wrong place to begin. You do not start with an agent that acts for you. You start by building your own vault, then a personal AI that works with it on your terms. Agents come after that, once there is a library worth acting on. For a business it is the same caution in a stricter form: agentic AI can work but only under real AI and data governance, not before.

It eventually leads to two twins rather than one. A personal twin lives with you, at home and on your phone. A professional twin stays at the office, owned by the company, with a firewall between them. The personal one never leaks into the company. The company one never follows you home. In the Apple TV series Severance, spoiler alert, a firm splits each worker into an office self and an outside self who never meet. It is a quiet horror, because the split is forced. What I am describing is the gentle inverse: you choose the wall, you own both sides and neither half is a prisoner.

Years ago I imagined everyone in the organisations I worked for keeping their own professional mind maps, all connected like synapses, sharing only what fitted. Unfortunately, it never happened. It can finally work now with AI twins running as agentic labour.

My grandfather kept a living library by hand and passed the method to me before either of us had a word for it. I am doing the same thing with better tools, because I would rather share what I find than keep it. He gave his library a method. I am giving mine a librarian.

I am not the first to walk this road. Dr. Thomas Roedl built ICOR and myPKA on the same foundation I keep returning to: plain text, any model you like, a folder that stays yours. Nate Herk teaches ordinary people how to build AI agents and automations, not just engineers. Their work is more built out than mine and worth your time. I am walking the same road in my own way with one emphasis: you should own all of it. Your notes now and one day the AI itself, running on your own computer.

This is the first AI Twin essay. The second, Field Report, is the record of the build itself. Read it when you want to see the system built milestone by milestone.

Come build alongside me. I am writing down what works and what breaks as I close the loop. I answer questions wherever I post this. No course, no product, no group to join. You do not need to be an engineer, though today it still takes patience and trial and error. Making that part disappear is the work I care about most.