mnemosyne-voice

Mnemosyne Voice

Mnemosyne OS runs local, offline text-to-speech engines. Some of them clone a voice from a short reference clip the human recorded themselves. This skill turns a script into a WAV file on disk and hands you its path — made for voice-overs a human will drop on a video timeline.

Nothing leaves the machine. No cloud service is involved.

You are borrowing someone's voice. That framing decides everything below.

Setup (if the voice tools are not available)

The tools (mnemosyne_voice_list, mnemosyne_voice_speak, mnemosyne_voice_status) come from the @mnemosyne_os/mcp server, and they are off by default. If they are missing, do not configure anything yourself — ask the human to add one line to their MCP config:

"env": { "MNEMO_VOICE": "1" }

and restart their agent session. The MCP server process is spawned once per session, so a session that started before the change keeps the old one.

Turning it on makes Mnemosyne ask the human to authorize voice:speak — a permission that is never auto-granted, not even to first-party apps, because its subject is their identity rather than their data. That dialog is the point. If it is declined, say so and stop; the memory tools keep working.

Three preconditions are not your fault when they fail, and each has a different answer:

Symptom	What it means	What to do
Cannot reach Mnemosyne	The app is not running	Ask the human to open Mnemosyne OS — the engines are Python sidecars inside it
ENGINE_NOT_INSTALLED	No local voice installed	Settings → Voice. It is a multi-GB download — tell them, do not wait on it
LICENSE_REQUIRED	Local neural TTS is a licensed feature	Nothing to retry. Say it plainly

The covenant — non-negotiable rules

1. A cloned voice is a person, not an asset. Produce only what the human asked for, in this conversation. Never make a voice say something the person did not choose to say — no impersonation, no words-in-their-mouth, however harmless the framing.
2. A clone name that does not exist is REFUSED, never substituted. On UNKNOWN_CLONE, ask the human which voice they meant. ⛔ Do not retry with a name you invented: a voice-over in the wrong voice sounds perfect and is worthless — the human may not notice until it is published.
3. Always report the file path and what was spoken. Audio the human cannot find is audio they cannot check.
4. You may list and choose voices. You may never create one. Recording a voice print is the human's own act, done in the app.
5. A render is a job — poll it, never restart it. The engine synthesizes one thing at a time, so re-issuing a render that is still running doubles a wait you are already in.