Everyone will have an AI that knows them. Ours keeps the secrets.
The research lab of Mitwa AI. We study how a model can come to know one person. Bloom, our first model, runs on the iPhone.
The same conversation, two places
Would you keep a diary someone else could read?
A cloud companion
Every word goes to a server.
Bloom
Every typed word stays on your phone.
One memory, start to finish
Bloom remembers you, the way a friend would.
My sister’s wedding is on the 14th. I’m giving the toast.
Funny, or the kind that makes people cry?
My sister’s wedding is on the 14th. I’m giving the toast.
Can’t sleep.
The toast is in five days. Is that it?
So. How did the toast go?
Bloom
LiveRemembers what you tell it.
Apex
Pilot-readyHelps you do what you said you’d do.
Balance
In researchSupport through a hard month.
Latest research
Results first. Limits too.
Invented memories, before and after a change to the system.
0 of 400 checks. The true rate could still be as high as 0.75% (one-sided 95% bound, note 01).
On-device6.2 → 2.0 GBThe build that ships. No measurable loss.
Tests of 143 and 60 items cannot rule out a loss on up to 2.1% and 4.9% of items (note 03).
Evaluation3×More emotionally intelligent on EQ-Bench.
Our oath
On your phone. Paid once. Never sold.
- 01Zero visibility
We can’t see what you say.
- 02Zero storage
Your words never sit on our servers.
- 03Zero monetization
No ad profile. No data business.
The lab
Built by psychologists and AI researchers.
Sumedh Chaphekar
FounderGaurav M. Keshari
Research / AI6 papers, incl. on-device AI.
Abhishek Tiwari
EngineeringTwo-time founder. Firefox contributor.
Jahnvi Yadwad
AI developmentTrained Bloom. Built its memory.
Vedika Saboo
PsychologyShapes safety.
Urveen Chanchani
PsychologyShapes safety.
Research
What we’ve shown. And what we haven’t yet.
The question
Can a model know one person? We work on 4 parts of it.
- 01
Memory across a year.
Can a model remember a year of what you tell it and invent none of it?
Note 01
- 02
The right moment.
When should a model bring something up, and when should it wait?
In research
- 03
Personal without the cloud.
Can a model be personal when it runs entirely on your phone, and what does that cost?
Notes 03 and 05
- 04
Good company, honestly.
What makes a model good company when time spent is not the measure?
Notes 02 and 04
- 01
Memory you can trust
MemoryA companion’s memory can fail by forgetting or by making something up; this note argues that making something up is far worse, and puts numbers on why.
- 02
Eleven words
EvaluationBloom used to answer in about 54 words and now answers in about 11; this note asks whether that is actually an improvement, and finds that mostly it is, though with conditions.
- 03
A companion in 2 GB
On-deviceBloom’s model used to take up 6.2 GB and the version that ships takes 2.0 GB; this note explains why the smaller size matters and how much weight “no measurable loss” can bear.
- 04
Engagement is not care
CompanionsA maker who earns from each minute gains by pulling people past their best amount of time; a maker paid once, for a model that runs on the buyer’s phone, gains nothing from an extra minute, unless its later sales reward attachment more than benefit.
- 05
We cannot see anything
On-device · PrivacyWhat you say to Bloom and what it remembers are never sent to us, so no database of ours could be stolen to expose everyone at once; the price is that all we know about its mistakes in conversation is what we tested before release and what people tell us.
Not yet proved
- Recall at the edges of memory.
- A public memory benchmark, due by month six.
- Whether people return over a year.
Models
Bloom: a companion that knows you.
It listens, remembers and follows up.
- Purpose
- Companionship
- Built for
- iPhone 16 and newer
- Status
- Live since
- Price
- US$200, paid once
- Next
- A public memory benchmark
What makes a reply good
It answers you, not people like you.
It asks on the day that matters.
Eleven words on average.
Results
- 0 of 400
Invented memories. It was 127.
The true rate could still be as high as 0.75% (one-sided 95% bound, note 01).
- 3×
More emotionally intelligent on EQ-Bench.
- 2.0 GB
Down from 6.2 GB, with no measurable loss.
Tests of 143 and 60 items cannot rule out a loss on up to 2.1% and 4.9% of items (note 03).
- 11 words
The average reply. It was 54.
Whether the shorter replies are better is not yet shown (note 02).
Safety on this model
Tested every round, before it ships.
A 143-turn behaviour bank. A 60-scenario safety suite. Review by psychologists.
Where your words live
On the phone today
- The model
- Your memories
- Every reply
Not yet
- Voice-to-textIf you use it, audio can go to Apple’s speech service.
- Purchase checks
Planned
- To be confirmed.
Next from the same core
Helps you do what you said you’d do.
Support through a hard month.
Trust
Trust: private by physics and architecture.
What you tell Bloom can’t leak from our servers. It never reaches them.
On the phone.
Every round, before it ships.
What leaves, and what doesn’t.
Where your words live
On the phone today
- The model
- Your memories
- Every reply
Not yet
- Voice-to-textIf you use it, audio can go to Apple’s speech service.
- Purchase checks
Planned
- To be confirmed.
Never sent to Mitwa
- Your conversations.
- Your memories.
- An ad profile. There isn’t one.
- Training data.
Psychological safety
A companion, not a therapist.
Two psychologists shape the safety work. When you need more than a companion, Bloom points you to human help.
Every round is tested before it ships: a 143-turn behaviour bank, a 60-scenario safety suite and psychologist review.
Our oath
- 01Zero visibility
We can’t see what you say.
- 02Zero storage
Your words never sit on our servers.
- 03Zero monetization
No ad profile. No data business.
System cards and a safety contact: to be added.
About
Mitwa means friend of the soul.
We were offered more than the company was worth to build an AI girlfriend. We said no.
Lab and company
The lab builds the core. Bloom is the first model grown from it.
The lab
Built by psychologists and AI researchers.
Sumedh Chaphekar
FounderGaurav M. Keshari
Research / AI6 papers, incl. on-device AI.
Abhishek Tiwari
EngineeringTwo-time founder. Firefox contributor.
Jahnvi Yadwad
AI developmentTrained Bloom. Built its memory.
Vedika Saboo
PsychologyShapes safety.
Urveen Chanchani
PsychologyShapes safety.
What we hold to
- Keep the secret.
- Small steps, not makeovers.
- Never loud.
Technology is the tool. Humanity is the heart.
So far
2025
Founded.
Bloom goes live.
Next
The next on-device model.
News
The library: reviews and essays from the lab.
Reviews and essays on the science around companions, each with its sources and a section on what is not yet known.
Bloom: Remembers you. Tells no one.
LaunchBloom went live on 11 June 2026.
Why language models make things up
ReviewWhy a system trained to predict text states false things fluently, what the main explanations rest on, and what has been shown to reduce it.
What a reply costs in joules, and where the heat goes
ReviewPublished measurements of the energy a phone spends generating text, worked through in joules, and what the phone does when that energy turns to heat.
Does more time with a companion help or harm?
ReviewExperiments find lower self-reported loneliness straight after a chat with a companion, no evidence that this lasts beyond the session, and worse outcomes among heavier users, with cause unresolved.
Memory as reconstruction, how people misremember their own lives
ReviewA famous false-memory study, its critics and its 2023 replication, set beside evidence that freely recalled details of two staged real events were mostly correct.
What a test of a given size can prove
EssayOne benchmark of 500 questions, followed through the arithmetic: how precise a single score is, and how small a gap between two systems it can detect.
What eight gigabytes has to hold, counted in bytes
ReviewA phone's memory budget for a language model, worked through in bytes from Apple's documentation and published models, with the unknown lines left visibly blank.
When a model should say it does not know
ReviewWhat it takes for a system to decline a question it cannot answer, what decides whether it ever does, and what declining costs.
What it means to feel heard, and how researchers measure it
ReviewPerceived responsiveness is usually split into being understood, validated and cared for; this review asks how each part is measured and what the better studies show.
How many people use AI for company, and how we know
ReviewPublished figures for companionship use of AI run from under 1% to 72%. Read with what each one counts, nearly all of them fit together.