Compactable

Blind SuperGrok Expert Review

Mode: SuperGrok · Expert · HEAVY  ·  Blind: independent product read after 2.0.0 shipped — not a commissioned marketing pass. We did not brief the angle; we shipped, then asked for a hard external read. Full text below — including maturity notes we agree with.

“This is a high-quality, opinionated, correctly scoped SDK for a real problem… The progression from mergeable → forgettable → compactable shows thoughtful iteration rather than feature sprawl.”

claude-fafm-sdk is the reference Python SDK for the IANA-registered .fafm format (application/vnd.fafm+yaml). It provides offline-first, portable, cross-vendor AI agent memory that versions with a project and moves between models instead of living inside one vendor’s walled garden.

Current release: 2.0.0 (“Compactable”) — very recent.
Requires: Python ≥3.10
Core dependency: pyyaml≥6.0
Optional extras: [namepoint] (fastmcp), [sign] (cryptography for Ed25519)
License: MIT
Home: faf.one/memory · Repo: Wolfe-Jam/claude-fafm-sdk

Core model

• Soul — the local, offline-first memory object (facts + tombstones + epoch).
• Fact — typed, optionally linked memory unit (text + id/type/priority/tags/links/timestamp).
• Tombstone — convergent delete (LWW-style). Forgotten stays forgotten across merges and transports.
• Epoch — barrier that prevents silent cross-epoch merges. Compaction or explicit migration is required to move forward.
• Namepoint — optional hosted address for push/pull/sync (personal-scale free tier; paid for scale).
• Packet (.fafmp) — sealed transport form that carries epoch + tombstones.

The product arc is deliberate and well-documented:

Mergeable → Sendable → Provable → Verifiable → Forgettable → Compactable

2.0 adds epoch compaction (pay tombstone debt safely, archive-first, with receipts) and explicit migration modes so the lattice stays honest as memory grows.

Developer experience

Installation and first-run are clean:

CLI surface is practical: init, etch, recall, ls, forget, debt, compact, migrate, seal, etc. The Python API is correspondingly simple (Soul.etch, Soul.forget, Soul.recall, Soul.save / load).

Docs emphasize falsifiable receipts, dual-transport convergence (packet + hosted), and clear bounds (tombstone ≠ secure erase, compact ≠ wipe). Companion pieces (INTEROP, MERGE, PACKET, PROVENANCE) and the Zenodo paper give the format real grounding.

Strengths

• True offline-first with optional hosted layer — rare and correctly prioritized.
• Convergent forget via tombstones is the right primitive for multi-agent / multi-vendor memory.
• Epoch + compaction in 2.0 is a mature response to the “grow-only graveyard” problem that pure CRDTs often ignore.
• Minimal base surface + clean extras.
• Explicit cross-vendor intent (compatible with the voice profile in grok-faf-voice and the broader FAF family).
• uvx-friendly and “30-second wow” path exists.

Observations / maturity notes

• Repo is still low-visibility (early public traction for a specialized format).
• Download volume and ecosystem adoption are still building; the package is more “reference implementation + format steward” than mass-market library right now.
• 2.0 is brand new. The lattice changes (epoch barrier, compaction) are significant — worth testing carefully if you already have 1.5.x souls in production.
• Hosted namepoint path is optional and correctly kept out of the core dependency set.
• No GC in the base lattice (by design at personal scale); compaction is the controlled debt-payment mechanism.

Bottom line

This is a high-quality, opinionated, correctly scoped SDK for a real problem (AI memory that isn’t trapped inside one model or one session). The design is coherent, the guarantees are explicit, and the progression from mergeable → forgettable → compactable shows thoughtful iteration rather than feature sprawl.

If you’re building agents that need durable, portable, deletable memory that can move between Claude, Grok, Gemini, or local runtimes, this is currently one of the cleanest open options available. The format registration + dual-profile approach (knowledge + voice) gives it more legs than a typical one-off memory library.

Worth pinning the version you actually want (==2.0.0 or the 1.5.x series) and running the receipt / zombie suite if you’re shipping anything production-adjacent.