Voice assistant · Running daily
ARIA
A voice assistant that remembers. Talks, recalls past conversations, pulls from a personal knowledge base, and hands real work to agents.
Why I built it
I wanted an assistant that behaved like a colleague, not a search box. That meant memory across sessions, a way to reach my own notes and documents, and the ability to delegate — none of which come out of the box with any consumer product.
What it does
- Cross-session memory
- She keeps a lightweight memory of what matters between conversations, without a round-trip to a model, so recall is instant and free.
- A second brain
- Her answers can draw on a personal knowledge graph — thousands of notes and documents — retrieved in the background so she never stalls mid-sentence.
- Delegation
- Ask for something that takes real work and she hands it to a separate agent, then pings your phone when it's done.
- Presentation mode
- She can drive a slide deck — images, notes, code — while she talks, receding to a presenter's corner.
- Noise-aware
- A half-duplex mode for outdoors or the car, so wind and road noise can't chop her replies.
aria2026
- ›session · memory loaded · 412 facts
- ›listen · vad ok · noise mode: off
- ›recall · 'the tax deadline conversation' · 2 hits
- ›retrieve · second brain · 3 docs · 140ms
- ›speak · streaming
- ›delegate · 'draft the follow-up' → agent
- ›notify · phone · done in 41s
- Real-time voice (WebRTC)
- LLM with reasoning
- Graph retrieval
- Agent delegation
- Self-hosted
What it taught me
The hard part was never the model. It was latency, memory, and making delegation feel natural — the same three things that decide whether AI actually sticks in a small business.