The Continuity Layer
The Continuity Layer · Standards, protocols and law

What exists, what is draft, what is in force

thecontinuitylayer.com/continuity/standards

Every instrument that bears on AI memory and continuity, with its date and its standing, and two columns that matter more than its name: what it can express about continuity, and what it cannot. Nothing on this page is a memory standard yet. That is the finding.

Checked 7 November 2026. Sources linked. Our own framing is labeled as ours.

Established The entries are the instruments themselves. The two columns are this site's reading.

Established standards

Vocabulary with a standards body behind it, older than the current wave, and reused rather than reinvented.

Standards-track publication· 30 April 2013

PROV-DM: The PROV Data Model

W3C Provenance Working Group · W3C Recommendation

Expresses
How to say where a thing came from: entities, activities, agents, derivation and attribution. The model behind every provenance claim on this site.
Cannot express, or is silent on
Nothing about memory, deletion or AI; it is the grammar, not the policy.

Obligations: Prove · Library entry

Published· 26 September 2025

Guidelines for Media Sanitization (NIST SP 800-88 Rev. 2)

Ramaswamy Chandramouli, Eric Hibbard · NIST · NIST Special Publication 800-88 Rev. 2

Expresses
Cryptographic erase: destroy the key, leave the ciphertext unreadable. The definition behind rung 8 of the erasure ladder.
Cannot express, or is silent on
Derived copies, indexes, or a model's weights.

Obligations: Forget · Library entry

Published· December 2023

ISO/IEC 42001:2023, AI management systems

ISO/IEC JTC 1/SC 42 · ISO/IEC 42001:2023

Expresses
A certifiable management system for AI, including records and change control.
Cannot express, or is silent on
The behavior of a memory system under deletion or poisoning; it certifies process.

Obligations: Govern, Prove · Library entry

Published· 26 January 2023

AI Risk Management Framework 1.0

NIST · NIST AI 100-1

Expresses
Govern, map, measure, manage: the US reference for AI risk practice.
Cannot express, or is silent on
Memory or state mechanics; voluntary and organizational.

Obligations: Govern, Prove · Library entry

Agent protocols with a home

Open specifications in use, now under neutral foundations. They standardize connection and delegation, not memory.

Published· 25 November 2024

Model Context Protocol (MCP) specification

Anthropic; now under the Agentic AI Foundation · Agentic AI Foundation (Linux Foundation) · Open specification

Expresses
How a model reaches tools, data and prompts; the way most memory servers are plugged in.
Cannot express, or is silent on
What a memory server must guarantee: no supersession, provenance, scoping or deletion semantics in the protocol.

Obligations: Carry · Library entry

Published· 9 April 2025

Agent2Agent (A2A) protocol

Google; now under the Linux Foundation · Open specification

Expresses
How agents find each other, delegate tasks and return results.
Cannot express, or is silent on
Delegation of state; audit and replay, dissent and voting are absent (Governance Gaps, 2026).

Obligations: Carry, Govern · Library entry

Published· 9 December 2025

Formation of the Agentic AI Foundation

The Linux Foundation · Linux Foundation press release

Expresses
A neutral home: MCP, goose and AGENTS.md at founding, with the largest labs as members.
Cannot express, or is silent on
Memory or continuity are not founding projects; interoperability of memory is left to other venues.

Obligations: Carry · Library entry

Memory interoperability, in draft

The first venues for making agent memory portable and verifiable. Nothing here has formal standing yet; the dates matter.

Draft, no formal standing· 19 June 2026

AI Agent Memory Interoperability Community Group, Charter v1.0

W3C Community Group (chair: Russell Jackson) · W3C Community Group charter (announced 16 July 2026)

Expresses
A charter to make agent memory portable and verifiable across vendors; use cases, an interoperability profile, conformance tests and a regulatory crosswalk as planned deliverables.
Cannot express, or is silent on
Community Group Reports are not W3C Recommendations; no deliverable published when checked; works one layer above a protocol it does not specify.

Obligations: Carry, Prove · Library entry

Draft, no formal standing· 11 October 2026

The Sovereign AI Horizontal Memory (SAIHM) Protocol

Russell Jackson · IETF Internet-Draft draft-saihm-memory-protocol-03 (Independent Submission Stream)

Expresses
Encrypted memory cells tied to a holder identity, a public audit log, revocable sharing, and a forget operation that issues a receipt and declares cryptographic or access-control erasure.
Cannot express, or is silent on
An individual Internet-Draft with no formal standing, not an IETF product; defines no export or migration mechanism.

Obligations: Govern, Prove, Forget, Carry · Library entry

Draft, no formal standing· 17 February 2026

NIST AI Agent Standards Initiative

NIST Center for AI Standards and Innovation (CAISI) · NIST; RFI docket NIST-2025-0035

Expresses
The US framing of agent standards: security, identity and authorization, agent-to-system interaction; an RFI closed and analyzed in 2026.
Cannot express, or is silent on
No finalized standard or guideline when checked; memory enters only through identity, authorization and logging.

Obligations: Govern, Prove · Library entry

Preprint· 28 April 2026

AI Identity: Standards, Gaps, and Research Directions for AI Agents

Takumi Otsuka, Kentaroh Toyoda, Alex Leung · AIFT, Singapore; Waseda University; Keio University · arXiv:2604.23280

Expresses
The map of identity mechanisms for agents (DIDs, Verifiable Credentials, SPIFFE and WIMSE, OAuth delegation, C2PA, Sigstore) and five structural gaps.
Cannot express, or is silent on
A report; multi-hop delegation and intent attestation are research directions, not standards.

Obligations: Carry, Govern · Library entry

Law and regulatory guidance

What is binding, where, since when, and what each is silent about.

In force· 12 July 2024

EU AI Act, Article 12: Record-keeping

European Parliament and Council · European Union · Regulation (EU) 2024/1689

Expresses
Automatic logging over a high-risk system's lifetime, for traceability. The record a deletion receipt would live in.
Cannot express, or is silent on
High-risk systems only, and later than planned: Annex III from 2 December 2027 and Annex I from 2 August 2028 under Regulation (EU) 2026/1744.

Obligations: Prove · Library entry

Published· December 2025

OWASP Top 10 for Agentic Applications 2026

OWASP GenAI Security Project · OWASP GenAI Security Project

Expresses
Memory and context poisoning named as a top-ten agentic risk (ASI06), with mitigations at the write path.
Cannot express, or is silent on
Guidance rather than a standard; mitigations are described as practice, not tested conformance.

Obligations: Govern, Forget · Library entry

Questions

Is there a standard for AI agent memory?
Not yet. As of November 2026 there is a W3C Community Group chartered in June 2026 to make agent memory portable and verifiable across vendors, with no deliverable published, and an individual IETF Internet-Draft (SAIHM, version 03 of 11 October 2026) with no formal standing. The protocols with a home, MCP and A2A under the Linux Foundation's Agentic AI Foundation, standardize connection and delegation, not memory semantics.
Does any law require AI systems to forget?
GDPR Article 17 requires erasure of personal data in defined cases, and the EDPB's Opinion 28/2024 holds that a trained model can itself contain personal data. Neither says what erasure means for a model, a summary or an embedding, and no regulator has prescribed a method. The EU AI Act's record-keeping duty (Article 12) applies to high-risk systems from 2 December 2027 for Annex III systems, after the 2026 Digital Omnibus.
Which of these does The Continuity Layer use?
The established ones for vocabulary: W3C PROV for provenance, SQL:2011 for bi-temporal data, NIST SP 800-88 Rev. 2 for cryptographic erase. The framework itself is this site's proposal and is labeled as such. Where a draft names a useful distinction, such as SAIHM's cryptographic versus access-control erasure, the framework adopts the distinction and credits the draft.