grok-faf-voice

Home: faf.one/voice Live demo: faf-voice.vercel.app

Voice agents that remember.

The Voice Memory Layer (VML) for Grok Voice. .fafm 🐘🎙️ LiveKit enabled.

Two profiles, one .fafm format. This is the voice profile (the Voice Memory Layer). For the knowledge profile — typed, cross-linked agent memory — see claude-fafm-sdk.

What's New in v0.4.0 — FastMCP 4

fastmcp floor moves to >=4.0.0 — the FAF Python family standard. No API change: the MCPaaS client pattern (Client + StreamableHttpTransport + except ToolError) is unchanged on 4.x, and grok's own httpx calls (xAI / ElevenLabs / transcribe) are untouched. pydantic>=2.12 follows FastMCP 4's floor. 196 tests green.

The Fast⚡️AF memory setup

pip install grok-faf-voice

from grok_faf_voice import VoiceAgent

VoiceAgent().run()

That's it. First run, your agent gets a namepoint — your @handle for FAF memory. Auto-generated free, no email needed. Want a branded handle tied to your email (or your X username)? Claim one anytime at mcpaas.live/voice/setup — one per valid email address.

What you need

• Python 3.10+
• An XAI_API_KEY — get one at x.ai/api

That's it. Namepoint and Voice key are provisioned for you.

Run it:

python my_bot.py console

console mode talks locally — no LiveKit cloud needed. Deploy to LiveKit later via python my_bot.py start.

How it works

• The agent listens via xAI realtime — five built-in voices: Ara · Eve · Leo · Rex · Sal
• Every session opens already remembering what was etched in past ones
• At session end, new memories consolidate silently
• Cross-session, cross-device, cross-model — your namepoint is the address

Voice swappable, memory permanent. ElevenLabs and Hume land in upcoming releases.

Custom Voices

xAI shipped Custom Voices on 2026-05-01. We shipped support 48h later.

from grok_faf_voice import CustomVoiceClient, VoiceAgent

# Clone a voice from a 90-120s WAV sample
cv = CustomVoiceClient()  # uses $XAI_API_KEY
voice = cv.create_voice("sample.wav", name="My Clone", language="en")

# Use the cloned voice in your agent — same two-line shape
VoiceAgent(voice=voice["voice_id"]).run()

Custom voice IDs (8-char lowercase alphanumeric, e.g. nlbqfwie) flow through VoiceAgent(voice=...) unchanged. The 30 free voices via console.x.ai work too — paste the voice_id and go.

Local souls (.fafm files)

Read and write souls straight off disk — no MCPaaS, no API key. Great for inspection, tests, backups, and loading souls written by other FAF-family tools (cross-vendor interop).

from grok_faf_voice import FAFMemory

# Load a .fafm soul from disk (any profile — voice or knowledge)
mem = FAFMemory.from_file("soul.fafm")
body   = await mem.get()                  # the soul body
recall = await mem.recall_for_prompt()    # ready for the prompt

# Write it back out (byte-identical roundtrip)
await mem.to_file("backup.fafm")

# Typed views — no manual parsing (v0.3.1)
mem.profile    # "voice" | "knowledge"
mem.facts      # parsed list of memory facts
mem.index      # top-level index (knowledge profile)

.fafm is the IANA-registered FAF Memory format (application/vnd.fafm+yaml). Validate documents against the published JSON Schema.

Want more?

Citation

If you use grok-faf-voice or the .fafm format in research or production, please cite the format paper. The Agents paper is the companion for the agentic era:

Wolfe, J. (2026). Permanent Memory and Instant Recall: The .fafm Standard for Multi-Profile AI Agent Memory. Zenodo. https://doi.org/10.5281/zenodo.20348942

Wolfe, J. (2026). Why Agents Need a Passport: .fafa — Portable Identity for the Agentic Era. Zenodo. https://doi.org/10.5281/zenodo.21951641

BibTeX

@article{wolfe2026fafm,
  title     = {Permanent Memory and Instant Recall: The .fafm Standard for Multi-Profile AI Agent Memory},
  author    = {Wolfe, James},
  year      = {2026},
  month     = {may},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20348942},
  url       = {https://doi.org/10.5281/zenodo.20348942}
}

@article{wolfe2026fafa,
  title     = {Why Agents Need a Passport: .fafa — Portable Identity for the Agentic Era},
  author    = {Wolfe, James},
  year      = {2026},
  month     = {aug},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21951641},
  url       = {https://doi.org/10.5281/zenodo.21951641}
}

DOI: 10.5281/zenodo.20348942 (.fafm) · 10.5281/zenodo.21951641 (.fafa)

We are the Open-Ended answer to Voice memory, and don't you Forget It. We won't.

.fafm 🐘🎙️ · application/vnd.fafm+yaml (IANA registered) · part of the FAF.one family · MIT licensed