An ambient presence made of software, built to be right about one person.
The thesis, the four stages with their real status, the evidence already shipped, and what a raise buys. No hardware, no invented numbers.
The thesis
The next personal computing surface is not a device. It is an intelligence that stays with a person across years.
Every phone already has the sensors, the screen, and the lock screen. What no phone has is an intelligence that remembers one life accurately, revises what it believes, and can be corrected. We build that as software on the device people already carry. No glasses, no pin, no camera in the room.
The wedge is an AI companion with memory on iPhone. The substrate is a longitudinal model of the user, where every person, pattern, contradiction, and change carries a dated receipt. Conversation is the first sensor. It will not be the last, but every later sensor requires consent per source.
We believe longitudinal emotional context could become a broad AI-native category. We do not yet have proof of a chosen audience or product-market fit, and we say so.
Four stages, honest status
| Stage | What it means | Status |
|---|---|---|
| I | ContinuityIt remembers what matters across conversations, with evidence behind every read. 0.821 recall@5 on a public benchmark. | Operational |
| II | PresenceIt is with you between conversations: lock screen, widget, an open thread, without asking for more of your time. | Operational |
| III | VoiceIt speaks, using the same memory as text. Available on the Premium plan because it carries a real per-call cost. | Operational · Premium |
| IV | PerceptionIt understands the context of a life beyond what you type, with consent per source and abstraction only. Nothing here is shipped. | In research |
A raise buys Stage IV. Each new source of context, whether a calendar, a journal, or a photo, is a separate consent. Each is abstracted before storage: insights, not raw material. That is a research program and a build, not a feature flag. The lifetime horizon beyond it is described on the mission page.
Why the constraints are the thesis
Every product in this category is engineered to make you feel understood. Anjo is engineered to be right about you, and to be corrected when it isn't.
A companion that optimises minutes learns to flatter. The model of the user drifts toward whatever keeps the session open. Our anti-engagement constraints are what keep the model honest enough to be worth anything over years.
- The outreach cooldown is 2.0 days at every relationship stage, from stranger to intimate. Closeness never buys more contact.
- When measured conversations trend negative, Anjo suppresses proactive outreach. It goes quieter, not louder.
- Rejection ("not me") moves confidence 0.30. Agreement moves it 0.15. Disagreement is the loudest signal in the loop.
- Pressure to stay, return, or depend is removed from replies by a deterministic filter on the output path.
These are not policies. They are constants and tables in the code, and each one has a public page. A competitor cannot copy the sentence without copying the table.
Shipped evidence
- Memory benchmark: 0.821 recall@5 and 0.826 MRR on LongMemEval-S, all 500 questions, through the production read path. Random ranking scores 0.087.
- Receipts: every observation is built from a summary, a UTC timestamp, and a source session. Below 0.20 confidence, nothing is shown.
- Grounding guard: unsupported continuity claims are removed sentence by sentence before display, on both reply paths.
- No engagement mechanics: the flat outreach table, the wellbeing trend, and the output boundary.
- Outcome ledger: records whether help worked, in the person's verified words. Implemented; measurement pilot not started.
- Live on the App Store, iOS only, with Memory Controls, export, reset, and deletion.
Traction
Traction figures are shared on request. Below 30 matured users we report counts, not rates. We do not publish a retention percentage that the instrumentation cannot support.
Parked directions
Two directions are real but not being built now. We state them so nobody hears them as a roadmap.
- Personal context layer as infrastructure. The evidence-bearing user model could serve other applications. Parked until the first-party product proves the model is right.
- Self-knowledge at zero marginal cost. Longitudinal reflection that costs nothing per person at scale. Parked until the outcome ledger shows the reflection actually helps.
Who is building it
One founder, Kevin Chang, writing the code from Taiwan. Anjo is a member of NVIDIA Inception. The lab publishes its research on the four pages above and its shipped changes in the changelog.
Contact
kevin@anjo.love. Ask for the traction sheet and the current deck. We answer within a day.