How it works
A context graph, a permission spine, and one rule: propose, never commit.
U.md isn’t a notes app with AI bolted on. Underneath is a governed layer that decides what is true about you, what agents may know and do, and leaves a trail of why.
Capture — as exhaust of the work you already do
Email, calendar, files, and past chats flow in as sourced nodes. You can also import a ChatGPT/Claude export, or just answer “what do you know about me?” Nothing requires you to sit and write a wiki.
Organize — into a typed, sourced graph
Each item becomes a node with a type, a source, and a confidence label — decisions and the reasoning behind them, not just preferences. Connections form between them. The graph compounds.
Retrieve — with citations, never guesses
Ask in plain language and get an answer that carries its sources and a confidence label. An unsourced “fact” is blocked before it ever reaches you.
Assemble — the Context Router
Before you prompt any model, the Router proposes the smallest-sufficient bundle of the right examples, methods, and decisions. You toggle what's included; it hands you a copy-ready block or shares it directly — every include audited.
Act — staged, then approved
U.md can take real actions in the world — re-date a CRM, draft a deliverable — but it stages them with the reason shown and waits for your tap. Approve, and it executes and records what it did.
The asset
Typed nodes carry provenance, confidence, and permission.
Every node knows where it came from, how sure it is (fact · decision · preference · inference), and who’s allowed to see it. That’s what makes recall sourced, the Router selective, and sharing safe.