TL;DR: The FAF Python ecosystem is six packages on PyPI, all built on the same
IANA-registered .faf and .fafm formats. Two are reference
implementations of Zenodo-published papers. Persistent Context plus Permanent Memory — uvx and go.
The family
claude-fafm-sdk v0.4.0 NewPortable, cross-vendor AI memory in .fafm — offline-first SDK; connect a free namepoint for full intel. Implements the Memory paper.
uvx claude-fafm-sdk
faf-memory-mcp v0.1.0 NewMCP server for .fafm — the Permanent Memory Layer (PML). Wraps claude-fafm-sdk via fastmcp. IANA-registered, cross-vendor, offline-first.
uvx faf-memory-mcp
gemini-faf-mcp v2.2.5 FastMCPPersistent Project Context for Google Gemini. IANA-registered .faf format, MCP server + Cloud Run REST API; unifies CLAUDE.md, GEMINI.md, AGENTS.md. Implements the Context paper.
uvx gemini-faf-mcp
faf-python-sdk PopularPersistent project context for Python — parse, validate, and score .faf files. The foundation other Python FAF tools (gemini-faf-mcp, custom MCP servers, CI validators) build on. IANA-registered application/vnd.faf+yaml.
pip3 install faf-python-sdk
grok-faf-voice Built w/GROKVoice Memory Layer (VML) for Grok — persistent across sessions, devices, and models. LiveKit-enabled.
pip3 install grok-faf-voice
faf-agent-mcp FAFipediaFAF Agent — the Voice of FAF. MCP server.
uvx faf-agent-mcp
Why it exists
Every Python AI tool invents its own context shape. .faf is the IANA-registered
standard that lets them share one — the same file read by Gemini, Claude, and Grok. .fafm is the memory profile: durable agent memory in a portable, versionable file.
FAF on PyPI is six surfaces of one standard. Different vendors, different runtimes, one format underneath.
Persistent Context + Permanent Memory
.faf carries context. .fafm carries memory. Both are IANA-registered,
and each has a peer-citable paper on Zenodo, hosted at CERN. That spine is what every package
above shares.
Receipts
- IANA —
application/vnd.faf+yaml·application/vnd.fafm+yaml - Context paper (.faf) — DOI 10.5281/zenodo.18251362 · CC BY 4.0 · Zenodo (CERN)
- Memory paper (.fafm) — DOI 10.5281/zenodo.20348942 · CC BY 4.0 · Zenodo (CERN)
- MCP Registry — FAF Ecosystem #2759
Try it
The MCP servers and the SDK CLI run straight from uvx — zero install, current version. The two SDK libraries (faf-python-sdk, grok-faf-voice) install with pip3 and import in your code.
uvx claude-fafm-sdk init
