The core layer inside Helm AI
Every device you own should already know you.
The Mirror is a personal relationship world model: one portable model of who you are, the people in your life, and the history between you, that any device can read from instead of starting over. Your phone, your glasses, and the robot in your kitchen should not have to meet you from zero every morning.
In development, no launch date. The memory core is built and running behind our current products. The general build is not released yet. We will tell you when there is something you can run.
How it travels
One person, three devices, no reintroductions.
Each device holds only what the task in front of it needs. The understanding itself stays in one place you control, and it follows you.
You speak
You say who your mother is, that she does not eat shellfish, that last Thursday you promised to call. The phone records it with its source: your own words.
It observes
Your glasses see the same colleague on the street. An observation arrives as pending evidence with a confidence value, never upgraded silently into a fact about you.
It acts
You ask it to host three colleagues on Saturday. It reads who these people are to you, what each of them cannot eat, and which of your plans still holds, then asks about the parts it is not allowed to assume.
What is actually in the model
Seven layers, not one profile.
A profile flattens you into an average. This does not. Every layer is separate, every claim carries its own evidence, and nothing becomes true about you without your confirmation.
YouThe person root
Long-term goals, current state, values, hard limits, the habits you actually decide by, and the boundaries of what any device may do on your behalf.
Your peopleRoles and relationships
Who they are, what you call them, the role they play this week, how close the relationship is, and when that changed.
Who decidesResponsibility and power
Who holds the decision, who carries the consequence, who controls the resources and the rules, and whether you are free to refuse. This layer is where a pattern of pressure becomes visible instead of normal.
Your historyTime and change
When something happened, whether it still holds, whether a preference is a long habit or a bad week, and how your commitments have shifted.
Where it came fromEvidence and confidence
Your words, an observation, or our own inference. Whether you confirmed it. Whether two sources disagree. How confident the system is, and why.
Right nowContext and intent
What you are actually trying to do behind the surface request, which people and history are relevant, and which memories must not be used this time.
Your permissionAuthorization and privacy
Whether this is really you, what may be called, what needs asking again, and how you view, correct, delete, and withdraw authorization.
The rule everything else follows from
You hold the helm. It acts in your interest.
Helm is the name of our company, and it is also the part that matters: the steering. This is not an assistant that decides for you, and it is not a device that quietly optimizes for someone else's revenue.
You are in command
Helm acts under your command, not the other way around. It thinks for you and acts for you, but you set the course. There is no version of this where the model holds the wheel.
It only acts in your interest
Within the bounds you grant, it can decide on its own. Every one of those calls has to serve you, never itself and never the platform. If a decision would only make us more engaged, we do not make it.
Nothing is confirmed for you
Every claim keeps its source and its status. An observation never becomes a trait of your personality because we were in a hurry. You confirm, correct, or delete.
Your data does not train anyone else's model
What you tell us stays yours. It is not used to train a general model that serves other people, and it is not sold.
It is portable by design
Your model is yours to take elsewhere. If you leave, you take the understanding with you, in a form you can read. No lock-in by design, because the whole point is that it is yours.
The door stays open for makers
The Mirror is built to be the layer that lets a device know whose preferences govern it. Individuals come first, but device makers and OEMs are welcome to get in touch about embedding it.
How you can get it
Run it yourself, or let us run it for you.
The local build is the point. Anyone should be able to run a person-centered agent on their own hardware, in their own interest, rather than a vendor's.
Local and self-hosted
Bring your own model and API key. Local models supported.
- Runs on your own hardware. Your world does not leave it.
- You supply the model and the key, including a fully local model.
- Planned for an open-source release, so it is not locked to us.
- Free for individuals, and it will stay free for individuals.
Hosted by Helm AI
For people who would rather not operate it themselves.
- Same model, same rules about your data, run for you at a cost.
- Pay for the compute, not for access to your own data.
- No engagement targets and no third-party advertising in it.
- Not launched yet. Pricing is not settled.
Get on the list
We will tell you when there is something to run.
No launch date, no waitlist queue, no drip campaign. One email when the local build is ready, and one when the hosted version exists. If you are a device maker or an OEM, mention that in the note and you will go to a human.
Questions
What people ask first.
Is this another assistant I have to open?
No. The Mirror is not an app you visit. It is the layer that other things read from, so a device you already use can behave like it has met you before.
Does it profile me the way ad targeting does?
The opposite intent. Every fact keeps a person root, a source, and a confidence. Nothing is silently promoted from observation to established truth, and you can see and delete the whole thing.
What happens when a device I own wants to use something private?
That is a decision you make, per device and per category. Authorization is part of the model, so a device that has not been given access cannot quietly reach for it.
Why would a device maker care about this?
Because the hard part of an assistant is not the model, it is knowing whose preferences govern the request. A fridge that understands who is in the house, and what they cannot eat, is a different product from one that does not. We are open to that conversation.
Can I take my model somewhere else later?
Yes, and that is deliberate. The local build is planned to be open source. Understanding you should not be something you can only get from us.