Private, always-onAI systembuilt around you

She holds the material itself: your sent mail, your voice notes, your documents, your meeting transcripts, held whole and dated. When you ask, the passage comes back as it was written, not as a summary of it. Nothing is compressed to make room, and nothing drops off the end when a conversation gets long. The difference is not the model. It is what the model has been given to read.

She will know. Everything you share gets captured as it arrives, so the voice note where you said it, or the message thread, or the meeting transcript, is already in your corpus, alongside your mail, your files and everything else you have handed over. When you ask, she retrieves the passage and tells you the date it came from. You get the sentence you actually said, not a guess about it.

Nothing starts from zero. Every document, message, voice note and correction is written into the corpus and stays there, so a conversation in November stands on everything since the interview. When you tell her she got something wrong, that correction is a fact she holds, not a preference that expires with the chat window.

Your email and calendar connect through your own accounts, with your own login. Day to day you reach her on Telegram, on a private line only you can write to. WhatsApp and Slack are not part of the standard build; we add either one on request and quote it separately. Beyond that: any file or document you drop in, voice notes, web pages and PDFs.

She reads back. Old exports, archives, previous chat histories, folders of documents you have carried between three laptops: hand them over and they are woven in with everything current. Most people find the first month is spent feeding her the past, and it is the part that pays off most.

It is read and filed into your corpus, and from that second it is retrievable by content rather than by where you put it. Your original file stays exactly as it is, and the text she keeps beside it is plain: it opens on any computer, in ten years, with or without us, and no part of it is locked inside a format only we can read. Nothing is ever deleted: superseded material is archived and stays searchable, so you can always see what a document said before it changed.

Obsidian gets one thing right and this keeps it: plain files, readable without any particular app, on a disk you own. Notion holds your work in its own store instead. The real difference is who does the work. Both are cabinets you tend: you write the note, you make the link, you keep the structure alive, and the week you stop, it stops. She does the tending. She reads what you send, files it and connects it to what is already there, and what she reads is not only notes you typed into an app: your mail, your calendar, your messages, the recordings and the documents all arrive in the same place.

She has read your own sent mail, so she is copying you rather than imitating a register. She also knows the file: last year's questions, what you told him in March, the numbers he is asking about. Drafting is where she starts on day one, and she gets closer to you the more you correct her.

She can read across years in one pass, which you cannot. Give her a standing instruction to watch something, a client going quiet, a project drifting, the third time a decision has been deferred, and she compares what is happening now against what happened the last four times. That is where most of the useful surprises come from.

She runs on standing instructions. You set up routines and monitoring, one at a time, and they run around the clock without you: the morning read, the watch on a project, the research that needs to happen every Monday. You add the next routine when you know what you want running.

She is built around one person. Your team can benefit from what she produces, but she answers to you: there are no seats or logins to hand out. She lives on a server, not on your laptop. She can also be wrong: she can misread a document, or read two things and join them badly. Tell her, and the correction becomes something she holds rather than something you have to repeat. And she cannot know what she has never been given, which is why the interview and the first weeks of history matter so much. Everything she is weak at on day one traces back to something she has not read yet.

On your own private server in a secure data centre, rented in your name. That rent is the server's own running cost and it is separate from anything you pay us. Not on our infrastructure and not on a shared platform. We hold access during the build, and at handover that access is removed and everything transfers to you, the server, the files, the keys and the assistant. After that there is nothing for you to administer: when you want us, for a change, a repair or a new routine, you give us access for that piece of work.

Isolation is the main answer: your own server, your own keys, no shared tenancy with other customers, and no copy of your corpus held by us. Which AI model does the thinking is your choice. For sensitive work we recommend self-hosted models, where the information never leaves your server. The connections to your email and calendar run through your own accounts and can be revoked by you at any time. At handover our access is removed entirely. The files stay plain and portable, so you are never locked in to keep them safe.

You sit for a structured interview about your life and your work, recorded and transcribed, and that becomes the foundation. Then the build happens without you. Connecting your accounts is four short steps of a few minutes each, followed by a setup call of sixty to ninety minutes. It is useful from the first week and compounds from there.

Almost none, because the only habit is capture. Forward the thread, drop the file, send the voice note on the way out of the meeting. There is no maintenance, no admin interface and no weekly tidy-up. The people who get the most out of it are the ones who send her things, not the ones who manage her.

Whitehay AI, the company Kormakur runs with Tom Graham. Kormakur built the original Edda for himself before anyone else had one. From your interview to a system you use every day is about two weeks. We take on a few builds a month, which is a deliberate limit: each system is built around one person's actual work rather than configured from a template.