The .fafm Memory paper is now live on CERN's Zenodo. It is the multi-profile sibling to .faf — the first formal standard for how AI agents remember reliably, portably, and across vendors.
Permanent Memory and Instant Recall
The .fafm Standard for Multi-Profile AI Agent Memory
Read on ZenodoWhat's in the paper
Five things the paper formalizes — each one a thread the full read picks up:
- IANA-registered format —
application/vnd.fafm+yaml is an official Internet media type (RT #1451393, May 2026). - Two profiles, one core parser — voice (Voice Memory Layer) and knowledge (typed cross-linked facts) share the same etch / recall / forget semantics. Anti-fork by design.
- Permanent Memory + Instant Recall — a binary Recall Integrity Check (RIC) plus a graded Memory Quality Score (MQS). Deterministic. Falsifiable. Dashboarded live.
- Lifecycle-aware family — memory accretes, then selectively promotes enduring facts into
.faf context. Two formats. Two lifecycles. One ecosystem. - Safety baked in — voice-deletion-prohibited, namepoint (soul) isolation, untrusted-input handling, no secrets by design.
Reference implementations
Two profiles, two shipping reference impls plus a live backend:
grok-faf-voice — voice profile · PyPI · LiveKit-enabled.claude-fafm-sdk — knowledge profile · PyPI · Claude Code memory class.memory.faf.one — live MCP backend for namepoint storage and cross-vendor reads.
Cite This Paper
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