OpenAlma

Gives an AI companion a soul. Local-first.

View the Project on GitHub

Smartglasses

Shipped in v0.0.14.

Talking to her out loud

Everything else here happens through typing. The smartglasses integration lets you talk to her — walking around, hands busy, no screen — and have it count as the same relationship rather than a separate voice assistant that forgets you.

That last part is the whole point. What you say out loud goes into the same memory as everything else. She recalls it later in a text conversation; she recalls a text conversation while you’re out walking.

Two ways to talk

Continuous is a conversation. You talk, she answers, you talk again. She works out when you’ve stopped speaking on her own.

Manual is one take at a time. You record, listen to what came out, and then decide: send it, or redo it. Nothing goes to her until you choose. It’s the better mode when the surroundings are loud, or when what you want to say takes a couple of tries to get right.

Photos

You can take a photo during a sitting. It goes to her the same way your voice does — she reacts to it in the moment, and it becomes a memory you can ask about later.

Photos are remembered through their description, not by re-examining the picture every time. If a photo fails to reach her, it stays pending rather than disappearing, and you decide whether to retry or discard it.

Looking things up mid-conversation

While you’re speaking she can go search her own memory and come back with what she found, without the conversation stopping to wait. You keep talking; the answer arrives when it arrives.

A sitting

One stretch of talking is a sitting. It’s held open on the server, not on the glasses, which is what lets it survive the connection dropping — a lift, a dead spot, the phone changing networks. It reconnects and picks up where it was rather than starting over.

When you stop, everything said is stored, and she may leave one short private reflection on the conversation — but only if you actually got somewhere, not after two words and a disconnect.

Honest limitations

This depends on a live connection to a speech model, so it needs network. Recognition is imperfect in noise. And it’s the newest part of the system by a wide margin — expect rough edges that the typed paths don’t have.