Saltar al contenido principal

Review, your agent shows you its work

The Review window is where your agent shows you a video or a picture it made and waits for your answer. It arrived in 2.1.0. Since 2.2.0 the agent can also make the video itself: your particle avatar, saying the lines you asked for, made on your computer.

The loop:

  1. The agent renders a video, for example an avatar film.
  2. It opens a review. The video shows in the Review window, and on your phone when you paired the review bot.
  3. You approve, ask for changes, reject, or leave notes on exact frames.
  4. The agent reads your answer, renders again, and opens the next version.

Connect your agent​

The agent talks to two MCP servers. Both come with the app.

  • The memory server, @mnemosyne_os/mcp, carries the three review tools.
  • The actions server ships only inside the app. It carries the avatar film.
  1. Open the connection panel

    Settings › Agents, section Connect my coding agent.

  2. Connect your agent

    Press Connect next to Cursor or Antigravity. For Claude Code, copy the two lines shown and run them. "Half connected: the actions server is missing" means the avatar film is not reachable yet.

  3. Restart the agent

    Restart it, then tell it: "Set up Mnemosyne OS for me."

The memory server alone can also be installed from npm. The details are on Connect your AI. Mnemosyne OS must be running for every tool on this page.

The avatar film​

Two tools of the actions server make the film.

ToolWhat it does
os_avatar_filmStarts a film: the app speaks each line, draws the face frame by frame, and encodes an MP4.
os_avatar_statusReads the state of a film by its job id: voice, frames, done or failed.

What the agent can set:

  • lines: 1 to 12 lines, 400 characters each, about 60 seconds in all. Each line can carry one expression: nod, emphasis, smile, browRaise, tender, curious, wink, surprise, ponder, joy, amused, sigh, relief.
  • face: one of 12 built-in faces: woman, man, claire, amara, kofi, mei, kenji, kitten, shepherd, labrador, bear, leo. Default woman.
  • aspect: 9:16 for phones, the default, or 16:9.
  • subtitles: the line at the bottom of the picture, on by default.
  • intro: brain, the particle brain that becomes the face, or none.
  • engine and clone: the voice. By default, the voice chosen in Settings › Voice.

What to expect:

  • The first time, a dialog in the app asks you to allow two permissions: avatar:film and voice:speak. They are sensitive and never granted on their own.
  • No credits are used. The work runs on your computer, about one second of work per second of film, after the voice.
  • The film lands in ~/.mnemosyne/generations/avatar-films/. The agent can open a review on that path directly.
  • A voice name that does not exist is refused, never replaced by another voice. With no voice chosen and none named, the film is refused with NO_VOICE.
  • The encoder is the one downloaded in Settings › Video. Without it the film is refused until you download it.
  • Your own avatars from the Avatar Studio are not reachable by the tool yet: it uses the 12 built-in faces.
  • macOS is refused for now: there is no encoder build for it.

The review tools​

ToolWhat it does
mnemosyne_review_openOpens a review on a video or a picture, with a title and an optional question. A retake passes the earlier review as parentId.
mnemosyne_review_waitWaits for your answers to one review. One call waits at most about 45 seconds; the agent calls again on STILL WAITING.
mnemosyne_review_getReads a review as it stands, without waiting.

The file must sit in ~/.mnemosyne/generations, ~/.mnemosyne/voiceovers or ~/.mnemosyne/outbox. Anything else is refused with OUTSIDE_SEND_FOLDERS. Accepted files: .mp4, .mov, .webm, .png, .jpg, .webp, .gif.

Answer in the Review window​

Review is not in the dock by default. Right-click the dock, choose Add a widget, then Review. Then ask your agent to use it. The list shows each film once, with its versions. When no review exists, the window says how an agent opens one.

  1. Watch the version

    Pick a film in the list. A retake shows Compare with the earlier version, side by side.

  2. Leave notes on frames

    Under Frame by frame, step with the arrow keys and press the up arrow to write a note on the frame shown. A note leaves as soon as you add it.

  3. Give a verdict

    Approve, Change or Reject, with text or a Voice note. Approve asks you to confirm first.

Keys in the player: Space plays, ← and → step one frame, Shift with an arrow steps one second.

What else the window tells you:

  • An agent is listening: what you send reaches it now. Or: No agent is listening right now: what you send waits here, and the agent reads it on its next call.
  • Notes alone are grouped. The agent wakes 30 seconds after your last note, once, with all of them. Cancel wake holds them; Wake the agent now sends them at once.
  • Send to the agent sends your text and notes without a verdict.
  • Send and tell the agent shows for a review opened by Claude Code. It leaves your answers in that session's mailbox, read when the session next stops or gets a message. A session already idle is not woken.
  • Folder opens the folder of an approved video, the file selected.

The list folds into a narrow rail, filters by state, sorts by state or date, and keeps favourites and pinned films on top. Delete moves the versions and their voice notes to the system trash.

Answer from your phone​

The Setup view pairs a Telegram bot made for reviews. Create it with BotFather, paste its Bot token, then send the pairing code to the bot. This bot is separate from the assistant's bot in MnemoHermes.

A paired phone receives each review with three buttons. A reply to the video, a text or a voice message reaches the agent the same way as an answer in the window. A voice message is transcribed on your computer.

Publish an approved video​

Nothing is published without your approval, and approving never publishes.

  1. Host on Cloudinary: with your Cloudinary keys saved in Setup, the video gets a public link. Anyone with the URL can watch it.
  2. Put in Buffer: choose channels and press Create to make drafts. You finish each post in Buffer.

The Cloudinary keys are checked against Cloudinary when you save them. The cloud name is in lower case: it shows at the top of Cloudinary's API keys page.

See also​