Faf Memory

Reproducible methodology + scripts for the .fafm binary tier — the receipt behind the 400+× type-filter speedup vs grep on AI memory corpora.

Built 2026-05-13 — the same day application/vnd.fafm+yaml was IANA-registered. This is the first real-world implementation of the registered format applied to AI memory-corpus storage, measured at scale (492 .md files).

---

Headline

Claude's persistent memory corpus, compiled end-to-end and benched against the status-quo .md + grep baseline:

Tier	Size	Cold load (492 files)	Type-filter query (warm)
.md (status quo — grep on prose)	2,099 KB	80.6 ms	29.5 ms
.fafmbin.gz (compiled binary)	996 KB	49.4 ms	0.072 ms
Ratio	2.11× smaller	1.63× faster	412× faster

Numbers are rounded down to 400+× in headline copy (strategic-undersell — the receipt holds the actual 412×). Full methodology, hardware, sanitization notes, and per-stage results in RECEIPT.md.

---

Follow-up receipts

The methodology has been scaled and validated cross-vendor since this repo shipped:

• xai-faf-proof — Grok + Claude co-built. Same methodology, run on the Smithsonian Open Access corpus (9,175 records, CC0 public-domain) AND a fresh Claude memory corpus (674 records). Peak speedups: 436× Smithsonian (within 6% of the 412× measured here — methodology scales), 1,399× Claude memory. Reproducible + falsifiable. See section 14 of RECEIPT.md for the scaled numbers.

---

Reproduce in 30 seconds

The repo ships with a 10-file sanitized pilot corpus at every tier (pilot/md/, pilot/fafm/, pilot/bin/) so you can run the benchmarks without supplying your own data.

Setup

git clone https://github.com/Wolfe-Jam/faf-memory-proof.git
cd faf-memory-proof
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run

Each block is paste-and-go — no comments inside, so any shell works.

1. Grep baseline on the pilot:

python3 query_bench.py

2. .fafmbin tier on the pilot (the 412× lane):

python3 query_bench_binary.py

3. Compile your own .md → .fafm → .fafmbin:

python3 convert_md_to_fafm.py
python3 compile_to_binary.py

4. Full pipeline + bench on your own memory dir:

SRC_DIR=/path/to/your/memory python3 scale_up.py

All scripts honor environment variables for input/output paths — defaults point at the bundled pilot (pilot/md, pilot/fafm, pilot/bin). See each script's header.

One-liner — fresh-clone smoke test

git clone https://github.com/Wolfe-Jam/faf-memory-proof.git && cd faf-memory-proof && python3 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt && python3 query_bench.py

Notes

• Pilot vs full corpus. Pilot bench (10 files) shows ~200× type-filter speedup; the headline 412× is the full 492-file run — the structured tier's advantage scales with corpus size.
• macOS: the system python alias may not exist — python3 always works. Use the venv above to sidestep PEP 668 ("externally-managed-environment").
• Type vs substring. Type/date filters dominate; full-text substring search is grep's natural strength — by design.

Requirements: Python 3.11+, PyYAML. Pinned in requirements.txt.

---

What's in here

Path	What
RECEIPT.md	Full methodology, hardware, ratios, sanitization notes
scale_up.py	End-to-end pipeline + bench runner (the 492-file run)
convert_md_to_fafm.py	.md → .fafm (structured YAML)
compile_to_binary.py	.fafm → .fafmbin + .fafmbin.gz (binary tier)
query_bench.py	Grep baseline benchmark on .md
query_bench_binary.py	Type-filter benchmark on .fafmbin
pilot/md/	10 sanitized .md memory files (the pilot corpus)
pilot/fafm/	The same 10, transformed to .fafm
pilot/bin/	The same 10, compiled to .fafmbin + .fafmbin.gz

---

Format

.fafm — IANA-registered as application/vnd.fafm+yaml on 2026-05-13.

• Sibling of .faf (application/vnd.faf+yaml, IANA-registered 2025-10-30)
• Spec: Wolfe-Jam/faf · MEMORY-FORMAT.md
• Paper: Zenodo DOI 10.5281/zenodo.20348942

---

The FAF cluster (for context)

• faf-plugin — Claude Code plugin for .faf context (FCL)
• faf-memory (coming) — Claude Code plugin for .fafm Permanent Memory Layer (PML)
• This repo — the falsifiable receipt the memory plugin's perf claims rest on

---

License

MIT. See LICENSE.

Authored by wolfejam (James Wolfe), with Claude as session collaborator.