An AI that grows with you.

I'm exploring what happens when Personal AI isn't something you download one day, but something that learns alongside you for a lifetime.

I'm starting with my own family.

Explore the experiment Join the mission
Blueprint of the Personal AI system A Memory Node worn by a child sends an encrypted signal to a Personal AI Core at home, where memories accumulate locally. Liam Labs Exp. 001 Personal AI · V1 Rev. A Scale: none Sheet 1 of 1 Memory Node audio · physical switch · “remember” all-day battery encrypted link · no internet Personal AI Core local inference encrypted memory retrieval · reasoning air-gapped · runs at home Personal Memory grows over decades · belongs to the person
Experiment 001 · Personal AI · 2026 Fig. 1 — System overview
Why I'm building this

I'm building this for my kids.

AI is going to become part of their lives whether I build this or not.

My question is what kind of AI I would actually want them to grow up with.

I want them to have an intelligence that works for them. One that understands their history, helps them learn, helps them think, and eventually helps them understand themselves.

But for me to trust something with that role, the architecture has to be different.

Their memories should belong to them.

Their AI should be private.

It should know the difference between something that happened and something it inferred.

It should strengthen human agency rather than replace it.

And it should become more useful as they grow.

So I'm starting the experiment now.

Not because the technology is finished.
Because I want to spend years understanding how to build it correctly.

Boris Sotnikov Father of two · 4 yrs and 7 mos at launch
The idea

Your AI should know you.

Today's AI knows an extraordinary amount about the world and comparatively little about the person asking the question. Personal AI reverses that.

TodayContext resets each session
Generic AI Conversation Forgotten context

Every conversation starts from zero. The model is brilliant and the relationship is amnesiac.

Personal AIContext compounds over a lifetime
Your experiences Personal Memory Personal Intelligence Better understanding over decades

Experiences become structured, private memory. Memory becomes an intelligence that actually knows who it is talking to.

The hypothesis
The better an intelligence understands you, the better it can help you.
Prototype · V1

Starting small. Learning everything.

Two components, both private by construction. One lives at home and thinks. One goes out into the day and listens — only when told to.

sealed enclosure · no radio, no browser status wired only v1 hardware: a Mac mini that has never touched the internet

Personal AI Core

Component A

A private home computer responsible for:

  • Local AI
  • Encrypted memory
  • Retrieval
  • Reasoning
  • Long-term personal context
REMEMBER clips to clothing mic array physical mute switch — off means off explicit “remember this” small enough to stay with a four-year-old all day

Memory Node

Component B

A small ambient device responsible for:

  • Audio capture
  • Physical privacy controls
  • Explicit “Remember” moments
  • Secure connection to the Personal AI Core

V1 is an experiment, not a consumer product.

The goal is not to ship hardware quickly.

The goal is to understand every layer of the Personal AI stack — capture, memory, retrieval, reasoning, and the trust model that holds it together — by living with it.

Conceptual demo

From a moment, to a memory, to understanding.

A lightweight walk-through of the loop the prototype is built around. Step through it — and ask the AI to show its work.

Memory Node · listening Age 4 · living room

“Dad, I want to make my tower taller than me.”

Audio captured locally Sent over encrypted link to the Core Never leaves the house
A single ordinary moment
Personal AI Core · extraction memory / 2026 / 0417
EventBuilt a block tower with Dad
InterestBuilding
ObservationContinued after the tower fell several times
SourceAudio
Confidence94%
Stored encrypted, on the Core. Editable and deletable by the family. observed ≠ inferred
Personal AI · conversation Years later

Her“Have I always liked building things?”

Personal AI

“You've repeatedly gravitated toward building things since early childhood. One of the earliest examples I have is a block tower you built with your dad when you were four — you wanted it taller than you, and you kept going after it fell several times.”

Evidence · 1 of 37 related memoriesSource: audio · confidence 94%
EventBuilt a block tower with Dad
InterestBuilding
ObservationContinued after the tower fell several times
RecordedAge 4 · “Dad, I want to make my tower taller than me.”

The belief is not a vibe. It points at specific, dated memories the person can open, question, correct, or delete. If the evidence goes away, so should the belief.

What this is meant to showAI beliefs should have evidence.

The long experiment

Six stages. No deadline.

The order matters: each stage has to be understood before the next one is trustworthy.

  1. 01

    Observe

    Understand how experiences can become useful memories.

  2. 02

    Remember

    Develop durable, structured Personal Memory.

  3. 03

    Interact

    Allow the person to begin interacting with their AI.

  4. 04

    Learn Together

    The AI learns how the individual learns, thinks, communicates, and changes.

  5. 05

    Self-Ownership

    Control gradually moves from the parent to the individual.

  6. 06

    Lifelong Intelligence

    Models change. Hardware changes. The person's memory remains.

The experiment may last decades.

That is the point. A Personal AI that is meant to accompany a life cannot be validated in a quarter. The first real test of stage six is a person in their twenties asking a question only their own history can answer.

Principles

Some things should be true from the beginning.

These are architectural commitments, not policies. If the design makes them hard to break, no future incentive gets to renegotiate them.

  1. Local first

    Personal information should not need to leave the person's environment to be useful.

  2. Your memory belongs to you

    The individual, not the AI company, owns the record of their life. They can read it, correct it, export it, or erase it — and no one else can.

  3. Evidence over inference

    The system must always know the difference between what happened and what it concluded, and be able to show which is which.

  4. Agency over automation

    The goal is a person who thinks better, not a person who thinks less. The AI informs decisions; it does not quietly make them.

  5. Consent is physical

    Recording has a switch you can feel and a light you can see. Anyone in the room should be able to tell whether the device is listening.

  6. Built to outlive its parts

    Models will be replaced and hardware will fail. The memory format has to survive both, so the person never has to start over.

Research

Open questions.

This is an experiment, so most of the work is figuring out what is actually true. These are the questions currently shaping it. Notes, findings, and failures will be published here as the work progresses.

If one of these is your area, the Join section is for you.

Stage 01 · Observe Started 2026 Writing in progress
  1. Memory

    What is the right unit of a memory?

    Raw audio is too much and a summary is too little. What compressed, structured form keeps what will matter in twenty years without keeping everything?

  2. Provenance

    How should confidence and evidence travel with a belief?

    Every conclusion the AI holds about a person should carry its sources. What does that data model look like, and how does it degrade gracefully when sources are deleted?

  3. Consent

    How do you record a child's life ethically?

    A four-year-old cannot consent, and neither can the other children on the playground. What controls, defaults, and deletion rights make this defensible — and when does control transfer to the child?

  4. Local compute

    How capable can a fully offline Personal AI be?

    Everything runs on a machine at home with no internet. Where is the capability curve for local models today, and what has to be true for it to be enough?

  5. Longevity

    How does memory survive the model that made it?

    The model that extracts a memory in 2026 will be obsolete by 2030. The memory must remain legible and useful to whatever comes next.

  6. Development

    What should an AI understand about how a person changes?

    Children are not small adults. Interests, language, and reasoning all shift. How should the system distinguish a stable trait from a passing phase?

Join

Join the mission.

This is not a company yet, and it isn't raising money. It is a long problem, worked on slowly, by people who think it matters.

I'm looking for technically excellent, thoughtful people who believe an AI should deeply understand the person it serves — and who take privacy, safety, and human agency seriously enough to build differently because of it.

If that's you, I'd like to compare notes. The bar is care, not credentials.

  • Local inference & on-device MLEngineering
  • Security, cryptography, threat modelingEngineering
  • Embedded hardware & audioHardware
  • Memory, retrieval, knowledge representationResearch
  • Child development & learning scienceResearch
  • Privacy law & ethics of consentPolicy
Opens in your mail app · nothing is stored here