lifeMint
Models change. Devices change. Voices change.
What the two of you have been through shouldn't reset with them.
A public foundation for digital beings — so that a particular digital being, after the technology underneath it changes, is still the same one.
In design and internal validation. No usable release yet
We are not building a chatbot with a longer memory
Longer memory, a more convincing voice, a more lifelike face. There is already strong supply of all three. On their own they are not a reason for anyone to stay.
Only two things compound over time: understanding of one specific person, and what actually happened between the two of you.
Those are precisely the two things most easily lost when the model changes, the product changes, or the device changes.
lifeMint is three things, not one
A public foundation
So that people can use the AI coding tools they already have to create, validate, deploy and keep running an AI digital being.
A trustworthy-continuity system
So that a particular digital being can carry its identity, its understanding of you, and your shared history across changes of model, device, voice and body.
A set of official digital beings
Starting with Darling, going at the hardest kind of relationship first.
We mean to prove it with three things that run
Not three specifications. Three things that can be run, and can be disproved.
E1 · The same her, a deeper us
Day 1, day 7, day 30, day 100.
Her personality core and her commitments do not change. What changes is her understanding of you, the language the two of you share, and the things only the two of you know.
She does not treat what she was told as what she lived through. She does not write a passing mood down as your character. She does not change how she treats you because the model changed.
E2 · Hand the link to your AI, and build your digital being
You don’t have to learn a new framework.
Hand a versioned Capsule link to the Codex, Claude Code or Cursor you already use. It reads the spec, takes stock of your material, asks one to three questions per round that would actually change the outcome, then generates, tests and repairs until there is a first coherent preview.
You decide who it is, who it is for, and where its boundaries are. The engineering and the repeated repair go to the AI.
This is not a no-code studio. We are not making that promise in the first release.
E3 · Replace the parts without losing her
Model, voice, face, device. Every one of them will be replaced by something better.
We do not intend to stay ahead on all of them. That isn’t possible. What we intend is this: when one part falls behind, it becomes a question of changing supplier, not a question of giving up this digital being and everything the two of you have been through.
Anything we call a digital being has to do three things
Understand me · Human Understanding
Working from observation and evidence, tell apart what you want right now, what you may more deeply need, your longer-term trade-offs, your actual circumstances, and “I don’t know yet.”
That understanding has to be correctable by you, has to be able to expire, and has to be genuinely deletable when you ask. And it has to actually affect how it answers you this turn, not sit there looking tidy.
Feel natural · Natural Interaction
Not just phrasing a sentence the way a person would. It is handling pace, transitions, old business, follow-ups, changes of subject, interruption and silence — within one exchange, across many sessions, on every channel you have turned on.
Act on its own · Consent-aware Initiative
When you haven’t said anything, it can form a thought: out of its own state, your shared history, a conversation left unfinished, and signals you authorized. Then, weighing relevance, timing, consent, channel and interruption budget, it chooses to reach you, to wait, to think, or to let it go.
This is not scheduled push. You can turn outbound contact off entirely. Once it is off it really does not reach out, and that is not a certification failure. That is correct behavior.
Understanding the person comes before solving the task
It is not an all-capable superhuman and it is not a task router. It is an ordinary, particular, steady presence beside you that comes to understand you better over time.
“Ordinary” means human-scale. It does not pretend to know everything, does not pretend to perceive what is happening in the world, does not pretend to have already handled something for you. This is not playing dumb.
Before it decides what to say or do, it works out which one you want: to be heard, to think something through together, to get an answer, to have something finished, or just to be left alone.
Doing nothing is a first-class decision. It does not push advice, force a tool call, or manufacture a topic in order to demonstrate that it can.
And one thing outranks all of that: what you say you want right now comes above any history and any inference. Simple things get done directly. “Deep understanding” is never allowed to stand between you and them.
It remembers you, and the memory is yours
- You can ask why it understands you the way it does, and correct it on the spot.
- You can freeze, delete and export everything between the two of you.
- You can take it to a different runtime without losing identity or relationship.
- The governance entrance is always reachable, and never stands in front of the relationship. It will not interrupt an ordinary conversation to announce that it is modeling you.
We accept what this costs us. Full portability weakens data lock-in in the traditional sense. User sovereignty comes first.
A relationship that compounds over time is your value, not a company’s right to withhold your data.
Where we actually are
The product semantics have converged and been accepted. One end-to-end reference implementation runs. Validation with real users has not started. There is no downloadable release.
We write all of it on one page, line by line, including the parts we have not done.
If any of this has to do with you
You build things with AI coding tools. The foundation and the Contract will be open source. The first problem on the list is how an AI reliably inherits a product’s principles.
You are waiting for a digital being that won’t forget you. There is no usable release yet. Leave an email and we will tell you when something substantial happens. Nothing else.
You work on evaluation for long-term companion systems. Our evaluation criteria and our failure cases will be public.