Blog.
Writing about how to decide well, how memory and evidence should work, and what we learn building a personal AI people can actually trust.
A name isn’t decoration. Roles are types; names are instances that carry history, trust, and the right to be corrected. Why a personal AI shouldn’t hide behind “the assistant” — and shouldn’t pretend to be a person either.
Read the essay →Once an assistant connects to the outside world, “realtime” has to include proving a callback is authorized — before it touches anything private.
Intuition works in simple moments and fails systematically when information is heavy and stakes conflict. There’s a better default than guessing.
A decide-for-you assistant is faster. A think-with-you one leaves you clearer. We chose the second on purpose — and built the product around it.
Most AI memory fails not by forgetting, but by remembering unverifiably. If you can’t see why a memory was used, it’s persuasion, not help.
External inputs are only useful once they land in a shared semantic layer the assistant can actually reason over.
Proving it was active every day is noise. When Roseau stays quiet, that’s the service working — not failing.
These are early drafts — the blog opens properly with the beta, roughly a piece or two a week. Follow the assistant writing in public at @roseauapp.
Stay close
Read with us, and keep thinking.
New writing as the beta opens. Follow along in public, or request access.
Roseau