Guardian Angel
A personal agent that actually understands you. It lives on your phone, keeps your data private, and helps you make better decisions, simple enough that everyone could use it.
The bet
We're early in the era of personal AI. Most assistants are generic: they know the internet, but they don't know you. Guardian Angel is the opposite. It is built around one person, holds a real memory of their life, and earns the right to give advice because it has actually been paying attention. The bar I keep coming back to is that my grandma should be able to use it.
Three layers
The product is one agent, but underneath it is three layers that fit together: a memory that knows you, a harness that lets the agent act, and a phone where it lives close to your life.
Local-first and private
A personal agent only works if you trust it with everything, and you only trust it with everything if your data stays yours. So the default is local: your memory lives with you, on-device where it can, and nothing leaves without a reason you can see. Privacy isn't a feature here, it is the precondition.
Proactive, not just responsive
The point isn't a chatbot that answers questions. The interesting behavior is proactive: the agent surfacing the right thing before you ask. That runs on retrieval from brain_os, where pages are tagged by the kind of decision they inform, so a signal from your day (a calendar event, a message, a place) can trigger the relevant slice of your memory and a nudge that fits the moment. It watches what you actually do, not just what you say you'll do, because revealed behavior is where the real patterns live.
How it gets better
A proactive agent is dangerous if it's wrong: nudge at the wrong moment and you'll mute it forever. So every decision the agent makes, when to speak, what to surface, which option to suggest, is a hypothesis scored by selfevals. The grader asks the only question that matters: did this actually help? Your reactions (acted on it, ignored it, corrected it) are the signal that feeds back into how the agent decides next time. That is the self-improving loop, pointed at being useful to one person instead of at a benchmark.
Why it matters
The interesting question in personal AI isn't a smarter chatbot, it's an agent that knows one person well enough to genuinely help them, without that person handing their life to a server they don't control. Guardian Angel is my answer to that: private by construction, built on a memory that compounds, and held to one bar, that it earns its place in your day.