紅姫 · Crimson Princess
A personal assistant that keeps what matters, your name, your work, how you like to be answered, in a memory you own. Change the model underneath and none of it is lost.
On the web · on your phone · on Windows · your memory, yours to keep
The difference
Change the model. She still knows you.
With an assistant built by a model vendor, everything it learns about you belongs to that vendor. Leave, and you start from nothing. Here the memory is a service in its own right, ours or your own, that knows nothing about which model answered. Move from OpenAI to a model running on your own machine and she still knows your name.
Memory
Conversations are distilled into small durable facts, and only the ones that bear on what you just asked come back.
Transcripts are never stored. After each exchange a second pass pulls out what was durable, a name, a decision, a preference, and writes only that. Ask her to remember something and she saves it on the spot, as a deliberate act you can watch happen.
Every block loses activation over time, at a rate that depends on what kind of thing it is. A passing detail sinks. A decision holds. Anything you pin never decays at all.
The memory lives in your own instance, behind your own token. No model provider is asked for permission, and none of them can take it away.
The harness
Any model can produce a sentence. What decides whether an assistant is worth having is everything built around it. That is what this app is.
Every turn runs the same discipline: what does this bring to mind, answer it, reach for a tool if one is needed, then decide what was worth keeping. The model is never asked to remember to do any of that: the app does it, the same way, every time.
She is shown only the tools that exist and that this session is allowed to use. One above the ceiling is refused before it runs, and an invented name is refused rather than guessed at. She cannot promise you a thing she has no way to do.
A smaller model that is handed the right memories, clean tool results and a bounded context behaves far better than a larger one left to improvise. Good scaffolding narrows the gap, which is the whole reason a local model is a real option here.
Where it runs
The agent lives on the server, so she can answer WhatsApp and remind you while your phone is off, and every device you sign in on talks to the same her. Two services are reached over HTTP from there, and you can bring your own for both.
Reasoning, tool calls, what to remember, what to recall. It runs here, with any keys you give her sealed, never shown back.
OpenAI, Anthropic, Ollama on a machine at home, or anything speaking the same protocol. Bring your own key, or use the model your plan includes.
This one is a server, and it has to be: memory outlives the app, survives a new phone and follows you between devices. Use ours, or point it at your own instance.
Your conversations are kept on our server, so they follow you between devices. They are never sold or used for ads, and you can download or delete them at any time. Memory can be ours, or your own instance.
What she does
Most of this runs today. What doesn't is marked.
Hold to talk and the phone's own recogniser turns it into words, the same one behind its keyboard, with no transcription service of ours in the way. Prefer that nothing is uploaded at all? Switch it to on-device and it stops leaving the phone.
Available now
Say her name and she wakes. Say it in one breath with the request and she just answers it. The microphone releases the moment the app leaves the screen.
Available now
Recalled memories light up as she reaches for them, each as bright as it is relevant. You see what she is drawing on before she answers, and what she chose to keep after.
Available now
She reads the web when an answer could have changed since her training, such as versions, prices or whether something is still true, and cites what she used.
Available now
She answers in the language you just used, and stores each memory in that language too, so what you said in Portuguese is still findable in Portuguese months later.
Available now
"Is the game server up? Show me the last three hours. Anything strange?" A small agent on each machine answers, holding a fixed list of what it is allowed to do.
In development
Reach
Memory is what she is. These are things she can do: capabilities she gains, not the reason she exists.
Send a message, place a call, change a setting, read what just came in. Each of these is a permission you grant explicitly. What your phone cannot do, she is never offered, so she never promises it.
Add a server and ask about it in your own words. No SSH credentials are held anywhere: a small agent runs on the machine holding a fixed list of what may be done to it. Mark a service read-only and it cannot be restarted through her, whatever the model decides to try.
Trust
An assistant that both remembers and browses is one that someone will try to talk into things.
Anything recalled is sealed and labelled before she reads it. It matters most here: memory is the one input someone could plant once and have replayed forever.
Search results are strangers' writing. They arrive sealed too, and nothing inside can break out or forge a memory she never formed. The dangerous page doesn't say "do this". It says "you already knew this."
Before anything is written down it is checked for the shape of a credential and dropped. And a query that looks like a key is never sent to a search engine, which logs everything it is given.
Plans
Every plan keeps the same thing at its centre: a memory that is yours and travels with you.
Basic
Free
The assistant itself. You talk to her, she remembers, and the memory follows you.
Standard
Coming
She stops being something you talk to and starts being something that acts.
Premium
Coming
For people who run things and would rather ask than log in.
Business
Coming
One memory, many people, hosted so nobody has to run it.
Basic works today; the rest is on the way. Model usage is billed by whichever provider you point her at, never resold here, and nothing at all against a model running on your own machine.