Memory you can edit
Every belief as a card that says where it came from. Tap the source. Tap to forget, and it is gone.
Meet Averil.
She texts you first.
She reads your email and calendar, watches for what matters, and messages you before you knew to ask. Then she asks before she acts.
Went through your inbox. Three subscriptions still charge you, and the gym has not seen you since March. Two emails are a week past a reply, one from your accountant.
draft the replies. cancel the gym
Drafts are in your email, and the cancellation is written. Say go and it sends.
Each weekday morning I can run this check and text you only when something needs you. Want that?
yes
Every weekday, quiet when nothing needs you. One more thing: check in for your Denver flight opens tomorrow at 9. Handle it? Aisle seat, like last time.
yes. and find me a flight back thursday night
1) 7pm, aisle, $340
2) 9pm, $280
Reply 1 or 2, or tell me something else.
1
Booked. Seat 14C, $340 on the Amex, under your travel cap. The confirmation is in your email and filed in the app.
Design partners first. Every friend who joins with your link moves you up three places.
Three things, and she does them well.
Everything important can be done by texting her. The app explains the thread; it never competes with it.
01
She lives in your messages
SMS or the app thread. No new habit to build. Replies land in seconds, short, matched to how you write, willing to disagree with you.
02
She notices things
A flight that opens for check in. A subscription that started charging you. The email that has waited three days. Every unprompted message spends from a hard budget, so silence is the default.
03
She asks before she acts
Autonomy is earned one rung at a time, per capability. You move any of them up or down. The model cannot.
- Read and brief
- Summaries, briefings, watching. No writes anywhere.
- Draft and ask
- She writes the email or fills the form, shows you the exact thing, and waits.
- Act and report
- Routine work inside a budget you named. Anything unusual becomes a question.
- Standing autonomy
- Named routines run on their own, each with its own cap and monthly summary.
- Never
- New payees, anything over your cap, deleting data, or changing the ladder itself.
The app is the glass box.
Not a second chat window. Everything she knows, everything she did, and everything she is about to do, in plain sight.
Every action, with receipts
What woke her, what she read, what she did, what it cost. The feed renders exactly what she read.
Automations with kill switches
Standing watchers and schedules as cards. What fires it, when it last ran, and a switch that stops it.
Her own browser, signed in as you
When she needs a site with no API, she texts you a link, you sign in with your own hands, and the session stays on her machine, not yours.
Priced where the gap is.
Between a chat subscription and an agent that bills like a contractor. We never retroactively paywall a habit.
Free
$0
- Reactive chat
- The daily brief
- One watcher
- One held draft, so you feel it once
Pro
$29 / month
- Full proactivity
- Email, calendar, memory
- Browser actions, signed in as you
- 400 runs a month, and she tells you at 80 percent
Concierge
$149 / month
- Heavy automation
- Priority execution
- A larger pool
- People for the physical world
Security is architecture, not a promise.
Assume the prompt has already leaked. The walls are built where a prompt cannot reach them.
Nothing you connect trains anything
Email and calendar are read for the task and kept no longer than the task needs.
Delete everything, in one action
Memory, history, artifacts, tokens. Deletion is real deletion.
The model never touches money
Spend caps and the ladder are set in the app and enforced at the tool boundary, not in the prompt.
