How it works
A person is a base model
plus a context.
People overlap on almost everything ordinary. What separates them is the last slice: experience, convictions, taste, the information they live in. Give a frontier model that slice, assembled properly, and it decides the way you decide. Everything below is how the slice gets assembled and where it is kept.
-
01
Your context makes the model you
Goals, taste, the history of your decisions, the way you argue a point. It lives in your own knowledge base as plain markdown in git: your property, portable, readable without us.
-
02
One identity, many hands
A master agent runs a swarm of specialists across mail, calendar, tasks and meetings. You talk to one identity, not to a toolbox of bots.
-
03
Five minutes a day keeps it current
Short daily and weekly rituals feed context back. The twin gets closer to you every week instead of drifting away from who you were in March.
-
04
Any frontier model, no lock-in
The model underneath is a swappable engine. The person is the context, and the context is yours. Change the engine, keep the twin.
The parts, named
The base
Your knowledge base is a git repository of plain markdown on your own disk. Notes, people, decisions, plans, the journal. Frontmatter for the machine, wiki-links between files, and a graph built out of those links rather than out of a database nobody can read. Nothing in it is a Clemond file format. Open it in any editor, copy it to any disk, and it still means what it meant.
This is also the answer to the question underneath every other question: your twin has memory because your memory is a repository, not because a vendor is storing your conversations.
A session
When you give the twin something to do it opens a session, and a session is a git worktree - its own working copy of the base. Two pieces of work never edit the same file behind each other's back, and every session's changes arrive as commits with an author and a time. You can read what it did after the fact, which is a different kind of trust from being told it did well.
A swarm
One master identity, many specialists. Reading a long thread, comparing three quotes, checking what a contract leaves out - work that a person would do in sequence runs in parallel, and the total is the slowest worker rather than the sum. You talk to one identity, not to a toolbox of bots, because the specialists report to it and it reports to you.
The rituals
Five minutes a day, half an hour a week. What happened, what is next, what was decided and why. That is what keeps the context current - without it a twin drifts towards the person you were in March and stays there. The rituals are the part people skip and the part that decides whether the thing is useful in month three.
The engine
The model underneath is swappable, and it is meant to be. The identity lives in the base: the rules, the memory, the history of decisions, the skills learned from mistakes. Put a new frontier model under it and the twin is the same twin, with better judgement. This is also the honest limit of any claim we make about quality: we are not the model, we are what makes a model you.
What it will not do
Small and reversible, it does. Outward, irreversible or expensive - sending, signing, paying, publishing - it does the work up to the last step and hands that step back. Not a rule it is following. A step it does not have.
The full account of what is sealed, who can read it and what a key costs is on the security page.
Start the same afternoon.
macOS, Windows and Linux. One download, one identity, and an interview that becomes the first slice of context.