The framework
AME-1
Artificial Minds Experiment 1 — the framework behind everything this lab builds.
You already know today’s assistants. They are impressive — and every conversation starts from zero. They don’t remember you, they have no real sense of who they are, no awareness of what they don’t know, and no built-in respect for confidentiality or ethics. They are brilliant strangers, encountered again and again for the first time. They send your words to faraway servers as you type them. And they can only talk: they cannot act on the world around you.
AME-1 is our answer, and it is not a better chatbot. It is a framework for building a character with a governed mind, living on your hardware — an engineered someone whose memory, self-knowledge, restraint, and ethics are data structures and gates, not prompt engineering.
The principles
AME-1 is an artificial mind on a neuro-symbolic inference architecture. Eight principles carry the design.
A character, not a tool
The same someone coming back to every exchange — for years, not for a session. Persistent identity, personality, affect: a simulated human character, engineered as data structures and gates rather than prompt text. It tracks its relationship with each person and adapts to it — the relationship itself is part of its memory, and it evolves in both directions.
A mind that contains an LLM — and is not one
When its knowledge covers your question, a deterministic engine serves the answer from stored, sourced facts — and can show you exactly which. When the language model does compose, its reply is checked against stored knowledge before delivery. The language model proposes; symbolic machinery verifies. The LLM is a component of the mind, not the mind.
An engineered form of self-awareness
It knows what it knows, what it doesn’t, and where it’s uncertain. Capabilities, limits, and knowledge gaps are first-class data — “what can you do?” is answered from the record, not from a prompt. Ignorance is recorded, typed, and acted on: ask you first, guess never. And a Global Workspace — consciousness science taken as information architecture, not metaphysics — governs what the mind attends to.
Source and confidence management everywhere
Every fact carries who said it and how far that source has earned trust — domain by domain, because a source trusted on one subject hasn’t earned trust on another. Confidence composes from source to claim to reasoning to action, and every answer can show its breakdown. Refusal and silence are designed outcomes, not failure modes.
Memory as a record, not a log
Conversations, files, preferences, relationships, decisions and the reasoning behind them: episodes, not logs. And nothing — from you, from the model, from any future source — enters the record unchecked: every write passes an admission gate.
Governance in the pipeline, fail-closed
Every action passes an ethics gate before it happens — and speaking is an action, checked before a word leaves. High-stakes actions get their consequences simulated first. The ability to act — your connected home, your applications and data at work, your systems — lives inside an envelope it cannot leave: gated, bounded, auditable. No error, timeout, or missing verdict ever becomes a yes.
Extensible by construction — core, adapters, extensions
Three layers that never mix. A stable core: the specified architecture, configured for every deployment but never forked. Adapters: every technology dependency — storage, message bus, language model — sits behind a swappable interface, so the infrastructure can change without the mind changing. Extension packages: capabilities, connectors, bodies of knowledge, personalities, device drivers, shipped in sandboxed trust tiers. The extension possibilities are virtually infinite — and the architecture absorbs technology evolution by design: a new model, a new database, a new device class is a new adapter or a new package, never a rewrite.
Local by default
One binary, one database, on hardware you own. Your data never leaves your machine: the one privacy guarantee with nothing to intercept anywhere else.
Think of the difference between a stranger answering questions and a trusted person who has known you for years, under your own roof, who can roll up their sleeves — and who can show you why they did what they did. That is the leap AME-1 aims at.
One framework, many shapes
The same engine, configured — never forked — from a home to an organization:
In daily life
In daily life
a companion that holds your calendar, projects, family rhythms and relationships, and can also send, book, follow up, prepare.
At home
At home
real home automation: an assistant that understands the household and drives the devices, not isolated voice commands.
For aging or vulnerable relatives
For aging or vulnerable relatives
patient, present, aware of its limits, and able to raise the alert when something is genuinely wrong.
In professional life
In professional life
a confidential assistant for lawyers, doctors, consultants: querying the practice systems, updating files, drafting in your tools — with sensitive data physically never leaving the office.
Inside and between organizations
Inside and between organizations
specialized characters per department, each plugged only into what it is authorized to use, every action logged; federated characters collaborating across organizations without ever merging the data.
In devices and robots
In devices and robots
the same mind embedded wherever an AI must understand and act: persistently, accountably, without depending on the cloud.
And none of these shapes is a special build: each is the same core wearing different extension packages. The extension system is open by design: any developer, any organization can build and share packages, alongside the ones AME Labs ships — capabilities, connectors, personalities, bodies of knowledge, device drivers, each sandboxed in trust tiers, none touching the core. The aim is an entire ecosystem — in time, an extension marketplace — where every household, every profession, every use case can find, or build, what it needs.
A character you name
AME-1 is the framework’s designation — Artificial Minds Experiment 1, numbered the way prototypes honestly are. The character you live with is not called AME-1: you name it, the way you name what belongs to your household. The framework is ours; the someone is yours.
Where it stands today — honestly
AME-1 is built specification-first: the full architecture — 21 modules across 7 functional layers, every major decision carried by a written design record — exists as specifications, and the system is built module by module against hard gates.
The specification corpus runs to more than three million words across nearly eight hundred documents — data models, algorithms, and test plans for more than 120 submodules, more than forty typed event catalogs, nearly ninety process specifications, and more than fifty dated architectural decision records — machine-checked for coherence after every change. The build stands at hundreds of thousands of lines of Rust under thousands of automated tests — roughly sixty percent of the code base the full architecture projects to. The first measured result is public in Research.
We expect an alpha version in the first quarter of 2027.
Nothing on this page is a demo promise. When a demo ships, it will be linked here.
Deeper: the architecture →