An independent AI lab, one person, one instrument.

The lab

Anjo is an independent AI lab based in Taiwan. We build persistent intelligence: software that learns one person across years, revises what it believes, and acts only with permission. Our first product is Anjo, an AI companion with memory, on iPhone. Anjo remembers what matters, connects patterns across time, and helps you see how you're changing. We are a member of NVIDIA Inception. The lab is one founder, and we publish our research and our measurements in public.

Why one person built this

Kevin Chang founded Anjo and writes the code. There is no team behind the name yet, and we say so on every page that could imply otherwise.

We built it alone because the problem does not need a large team to start. It needs a working memory, a model of the user with evidence behind every read, and a correction loop. Those three mechanisms fit in one codebase. What they need is patience and honest measurement, not headcount.

One person also has one advantage a larger company gives up. We answer to nobody who wants the product to be used more. So we can build the branch an engagement business cannot afford: when conversations leave a person worse, Anjo goes quieter. That decision is a table in the code, not a policy memo. It is easier to keep when one person owns the table.

The cost of one person is real. Releases are slower. Review comes from outside tooling and from users, not from a second engineer down the hall. We would rather publish that limit than hide it.

The mission and the five laws

Build a new kind of intelligence that can carry the continuity of a human life, and put it within reach of every person on Earth.

An intelligence that stays in a life for years must keep promises a chatbot never has to. These are ours. Each one is a mechanism in the code, and the mission page links every law to the page that proves it.

1. It must remember. What matters carries into tomorrow.
2. It must be right, not flattering. Every read carries the moment it came from.
3. It must be correctable. Being told "not me" moves the model twice as far as agreement.
4. It must never want more of you. Closeness never buys more contact.
5. It must belong to you. Every memory layer can be seen, changed, hidden, or erased.

The full statement, the lineage, and the five-stage horizon with honest status are on /vision.

How we work

The research is public. Each mechanism we rely on has a page that names the file it lives in and the limit it does not cross.

• Memory benchmark: the memory scores 0.821 recall@5 on LongMemEval-S, all 500 questions, through the production retrieval path.
• Receipts: every person, pattern, contradiction, and change in your model carries a dated moment.
• Grounding guard: an unsupported "last time you said" is removed before it reaches your screen.
• No engagement mechanics: the outreach cooldown is flat at every relationship stage.

We publish the benchmark rather than describing it. The number comes from a stored baseline file, and a test fails if the website drifts from it. The weakest slice, knowledge updates at 0.756, is on the page for the same reason.

We are also building an outcome ledger. It records whether a piece of help produced the result a person wanted, in their own words, verified against what they typed. It is implemented, and its two-week measurement pilot has not started. We will publish what it finds either way. The write-up is in lab notes.

When a mechanism is measured, we show the measurement. When it is not yet measured, we say that instead.

The Honesty Charter

What Anjo will and will not say about itself is written down and enforced on the output path, not left to the model's mood. Anjo does not claim to be human, does not pressure you to stay, and is not therapy, medical care, or crisis support. Read the Honesty Charter.

Contact

Write to Kevin directly at kevin@anjo.love. Press and investor pages are at /press and /investors. For account help, use /support.

Anjo is artificially generated AI, not a human. It is for adults 18 and over. Companion chatbots may not be suitable for some minors.