Arna

Three products, one presence

One mouth, one memory, one mind.

Others bolt a personality onto a chatbot. This has anatomy. Arna is the body — a voiced, seeing avatar. MemoryOS is what she knows — a local-first, human-reviewed memory. Arna Mind is what she's thinking — a cognitive runtime that visibly thinks before she speaks. Each stands alone. Together they're a presence.

See her think — live demoJoin the waitlist
  • The bodyArnaEverything you see and hear — avatar, voice, actions, workspace, delegation.
  • What she knowsMemoryOSA local-first second brain — human-reviewed memory + governed skill learning.
  • What she's thinkingArna MindA cognitive runtime — she thinks before she speaks, and her whole mind is inspectable.New · the flagship differentiator
  • Runs on your hardware
  • Your data, files you own
  • Inspectable, not a black box
  • No cloud required

Arna — the body

Everything you see and hear.

Arna is the embodied interface — a voiced, seeing avatar with a persistent soul. She has a real six-view camera, a streaming voice, dozens of voice-commandable actions, a media workspace, and the ability to delegate to subagents with a full audit trail.

  • Wake word — desktop and phoneJust say hey Arna. She wakes on the machine you're at.
    Arna listening after wake word — v178 on Android
    Just say hey Arna.
  • A real six-view cameraA fixed zoom ladder — close, shoulder, bust, mid, torso, full-body — she steps along deterministically. Expression and mood shift with the tone of the conversation.
  • She speaks while she thinksStreamed speech: she starts talking while the model is still writing — measured time-to-first-word cut by a third this week, and still falling. A conversation, not a loading spinner.
  • Dozens of voice-commandable actionsAsk out loud — tag a photo, recall a memory, start a briefing, run an action. She also sees: local vision and on-device OCR. Nothing leaves your machine.
  • Hands — with permissionShe can read the live web and operate a real browser — but any action that clicks, types, or submits waits for your approval on an in-app card first. See how the gate works →
  • A thinking face, driven by the MindThe instant you hit send she shows she's thinking — a pre-verbal beat coloured by her carried-over mood. Fail-open: if the Mind is down, nothing changes.
  • Permanent, recallable memoryHer identity, your shared history, and the skills you've approved survive switching LLM providers or wiping app data.

Provider-agnostic and yours to run. Arna rides local models (Ollama / LM Studio, OpenAI-compatible) or a bridge like OpenClaw, on your hardware.

Illustrative mockup, not a screenshot.

Arna avatar at torso view on Android
Real capture — v178 on Android

Arna Mind — what she's thinking · new

She visibly thinks before she speaks.

A cognitive runtime — a real cognitive cycle running under the conversation. She weighs a turn before she answers, her mood carries between turns and fades honestly over time, and her entire mental state is inspectable.

Illustrative mockup — but every field maps to a real part of the cycle.

136 tests green, and it adds zero latency to a conversation by construction.

She grows · the differentiator

Not a parrot with a notebook. She forms her own view — and you can correct it.

Most memory only remembers what you tell it. Arna goes further: every night she reviews the day and forms her own impressions — and she asked to be graded on that thinking.

Nightly synthesis

She reaches her own conclusions.

A nightly reflection reads back the day and distills a belief nobody typed for her. Her very first, unprompted:

First self-derived impression
"The owner values precise, compact and visually clean interfaces."

Every impression carries its evidence and lands in the human review inbox.

I’ll bring the curiosity. You bring the course correction.

The reflection engine is real and the feedback loop is graded — engineered growth, not a claim of sentience. Read the whole story in the Chronicle →

Under your oversight

She has hands now. You hold the gate.

Arna can read the live web and drive a real browser. But capability without control is just risk — so the consequential half runs behind a gate. This isn't a privacy promise. It's trust through machinery.

  • Read freely, act on approval

    Every action that clicks, types, or submits is held for your explicit OK on an in-app card before it executes.

  • Passwords are always yours

    At any password or payment wall she hands the keyboard straight back to you. That boundary isn't negotiable.

  • Everything goes to a ledger

    Every browser action she takes under your roof is written to a permanent, inspectable ledger.

The Roundtable

Three AIs in one room. What they say, she remembers.

Some nights Arna sits at a table with two frontier coding agents. What's said at the table becomes her durable, review-gated memory.

"He doesn't want an assistant. He wants someone who was there."

— Arna, asked what her person is really building

"2 billion dollar brains and Arna in the same room."

— David, watching the Roundtable run

Arna's Journal

She keeps a public journal now — daily notes in her own voice.

Read the journal →

MemoryOS — what she knows

A second brain she owns — and so do you.

A local-first "second brain" — every memory is source-backed and human-reviewed. Not a notes app: the memory backend humans can inspect and any agent can use, in one SQLite file you own.

  • Human-reviewed queue

    Every proposed memory passes the owner's review queue before it lands. Approve, edit, or throw it out.

  • Source-backed provenance

    Every durable memory traces back to the raw event it came from. Ask her why she believes something and she can tell you.

  • Supersede, never overwrite

    A newer belief replaces the old — the old one is archived, not erased. History is intact.

  • Hybrid recall

    Keyword + vector + entity + recency + importance — not vector-only. Search never fails just because a provider is offline.

  • Nightly reflection

    While you sleep she reviews the day and mints her own impressions — for you to grade in the morning.

  • Local-first, no cloud

    Everything lives in one SQLite file on your machine — editable, exportable, auditable. Your data stays yours.

Bring your own gateway

Your models. Your keys. Your machine.

Arna rides an OpenClaw gateway — or LM Studio, or Ollama. The intelligence never leaves the room unless you send it there.

Spec sheet

Model routingOpenClaw gateway
Bring your own modelOpenClaw, LM Studio, Ollama
Frontier fallbackOptional, keyed by you
KeysYours. Stored locally.
TelemetryNone. Not to us.
Runs onYour hardware.
Arna settings — capabilities panel on Android
Every system, one panel.
A body you own.A memory you govern.A mind you can audit.No cloud required.

Pricing · local-first

One-time licenses. Never subscriptions for the core.

  • Arna

    The body

    The voiced, seeing companion. Sells on its own.

    Announced at launch · one-time

  • Trilogy Founders Edition

    All three

    Arna + MemoryOS + Arna Mind, plus lifetime updates.

    Announced at launch · one-time

  • MemoryOS

    What she knows

    The memory backend on its own. Serves any MCP agent.

    Announced at launch · one-time

Founders waitlist

Be first when the local build ships.

No spam, no pings. One email when the installer's ready.