Permanent Memory and Instant Recall: The .fafm Standard for Multi-Profile AI Agent Memory

We introduce the .fafm standard (application/vnd.fafm+yaml), the IANA-registered media type for AI agent persistent memory and the sibling format to the .faf context standard. The format is multi-profile by design — a single core parser supports both the original Voice Memory Layer (VML) and a typed knowledge profile for agent runtimes — yet retains a single etch / recall / forget semantic surface, full forward and backward compatibility, and an explicit anti-fork architecture. We pair the format with two architectural commitments: Permanent Memory, the guarantee that declared memory must always recall (a binary integrity property), and Instant Recall, the graded quality of what was recalled (a slot-weighted score). The Recall-Gate / Quality-Grade scoring model derived from this pair distinguishes persistent-memory formats from static-context formats, is fully deterministic, and is published as a live dashboard rather than a frozen claim. Empirical validation runs against the Wolfejam Test & Technical Compliance (WJTTC) corpus across both profiles, with cross-application coverage spanning voice agents (LiveKit deployments, Grok Voice SDKs) and knowledge-runtime adapters (Claude Code memory class). Recognition signals at publication time include IANA registration, MIT-licensed open-source distribution, and an established install base inherited from the prior .faf ecosystem — which itself produces a measurable secondary observation: format-sequential adoption velocity, the empirical finding that an IANA-registered format published second into an established ecosystem propagates materially faster than the foundational first format.