Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
faf-memory-mcp
MCP server for `.fafm` — the Permanent Memory Layer (PML) for Claude Code, Cursor, Grok, and any MCP host.
Cross-vendor persistent memory in a file you can read. Offline-first. Your soul, your bytes.
Wraps claude-fafm-sdk via fastmcp. Receipt: 400+× faster type-filter queries vs `grep` on a real 492-file AI memory corpus — falsifiable methodology at Wolfe-Jam/faf-memory-proof.
Why
Every AI session, your agent starts blank — no memory of what you decided yesterday, what worked, what was tried. The fix isn't another hosted memory service; it's a file the AI reads at session start and writes back to as it learns. That file is .fafm — IANA-registered, cross-vendor, and 996 KB / 49 ms cold-load for a 492-fact corpus (full numbers: the receipt).
Install
uvx faf-memory-mcpUse in Claude Code (recommended)
Via the faf-memory plugin (coming to claude-plugins-community):
/plugin install faf-memoryOr wire .mcp.json directly:
{
"faf-memory": {
"command": "uvx",
"args": ["faf-memory-mcp"]
}
}Tools
| Tool | What it does |
|---|---|
etch(text, id?, type?, priority?, tags?) | Write a durable fact. O(1) dedup by id. |
recall(query?, type?, tags?, min_priority?, limit?) | Filter (substring + type + tags + priority floor) → rank by priority then recency. |
list_facts() | List ALL facts — no filter, no rank. |
save_soul(path?) | Persist the soul to a .fafm file. |
load_soul(path?) | Load a .fafm from disk, replacing the in-memory soul. |
Configuration
| Env var | Default | What |
|---|---|---|
FAF_SOUL_NAMEPOINT | @local | Initial soul identifier |
FAF_SOUL_PROFILE | knowledge | .fafm profile (knowledge or voice) |
FAF_SOUL_PATH | soul.fafm | Default save/load path |
Format
.fafm is IANA-registered as application/vnd.fafm+yaml (registered 2026-05-13). Sibling of .faf (application/vnd.faf+yaml, 2025-10-30).
- Spec:
Wolfe-Jam/faf· MEMORY-FORMAT.md - Paper: Zenodo DOI 10.5281/zenodo.20348942
- Reference impls:
claude-fafm-sdk(Python) ·grok-faf-voice(voice profile, Python) - Cross-vendor proof: a
.fafmwritten byclaude-fafm-sdkreads cleanly viagrok-faf-voice, both directions tested.
How to know it's working
- Your agent doesn't ask "what is this project?" twice in the same week.
recall("X")surfaces what you etched last session, not just this one.- The
.fafmfile grows readably — it's plain YAML, diff it like code. - Open the same
.fafmingrok-faf-voice— same facts. Cross-vendor proven.
Tradeoff note
Biases toward deterministic recall (substring + type + tags + priority + recency) over semantic recall. For semantic/ranked recall + LLM smart-merge, see hosted namepoints in claude-fafm-sdk. Offline-first ≠ offline-only.
The FAF Memory cluster
- This repo — `faf-memory-mcp` — the MCP server (wraps the SDK)
Wolfe-Jam/faf-memory— the Claude Code plugin that installs this MCPWolfe-Jam/faf-memory-proof— the falsifiable receipt (412× methodology + scripts)claude-fafm-sdk— the open Python SDK this server wrapsWolfe-Jam/faf— the format spec + IANA registration
License
MIT.