grok-faf-voice
Home: faf.one/voice
Live demo: faf-voice.vercel.app
⭐ Star it on GitHub: bookmarks it for you, helps other devs find it too.
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 v0.5.0 — The Ask-First Memory Edition
VoiceAgent() opens on the memory stored at your namepoint, and only writes the call back when you pass etch=True. 202 tests.
- Ask first. Without
etch=True the session opens on what is stored and the tool list has no etch_memory. etch=True brings back the etch log, the etch / recall / merge tools and auto-merge.
- Pinned model. The realtime model is
grok-voice-think-fast-2.0.
- Full notes: CHANGELOG.
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
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
