Is a continuity layer the same as AI memory?

No. Memory stores the past; a continuity layer governs what of it is still true, who may change it, why it changed, and whether it should be carried forward at all, and it can forget. Memory is necessary for continuity and not sufficient. Products like Mem0, Zep, Letta and the assistants' own memory features are memory systems; this site grades them on the questions a continuity layer would ask.

Why is forgetting part of continuity?

Because a record that cannot be corrected, revoked, expired, deleted and verified gone is a liability, not a working relationship. The 2026 research shows that deleting a stored memory leaves its influence in summaries, plans and indexes, and that deleted facts return within a few sessions through checkpoints, copies and consolidation. The erasure ladder names twelve operations so a “delete” claim can be read for what it guarantees.

Is the Continuity Framework an industry standard?

No. It is this site's proposal, versioned and signed, and every block of it is labeled as established, emerging, ours or a hypothesis. The established vocabulary it rests on comes from the sources in the research library: W3C PROV for provenance, SQL:2011 for bi-temporal data, NIST SP 800-88 for cryptographic erase, and the 2026 surveys for lifecycle and governance.

The Continuity Layer

AI doesn’t need more memory. It needs continuity.

thecontinuitylayer.com/continuity

The idea this site is named for: what an AI system has to carry forward, under what rules, with what evidence, and what it has to be able to let go. Read it, cite it, or use it to judge the products in the directory.

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

Frontier models reason well. What they lack is a present that survives: give one the same project two weeks apart and it starts from zero twice. Memory features patch that with saved facts, longer windows and background summaries, and the patches hold facts, not the working understanding of a relationship. A continuity layer is the part of a system that holds that understanding: what is true now, what changed and why, by whose authority, with what evidence, across every session, tool, model and host, and with the ability to forget.

The idea in one screen

Agents, models, tools, surfaces
Execution. Each is handed the same governed state.
↑ same state, every surface ↓
The continuity layer
What is true, what changed, why, by whose authority, with what evidence; and what has been let go.
Carry · Govern · Evolve · Prove · Forget
↑ reads, never trusts blindly
Memory and retrieval
Storage. Vector stores, graphs, files, a model's weights.

A continuity layer is the part of an AI system that carries forward governed state across sessions, tools, models, and hosts: what is currently true, what changed, why it changed, by whose authority, and which evidence supports it. It sits above memory and retrieval, and below agent execution, turning stored history into usable, auditable continuity; and it can correct, supersede, revoke, expire, and forget what it carries, with evidence that it did.
The Continuity Layer, working definition v2 (7 November 2026)

The definition is this site's own and is versioned: v1 (July 2026) is kept on the definition page. The thesis behind it, in eight beats with a published SHA-256, is at the thesis. The plain-language version, for someone who has only felt the problem, is Why AI forgets.

Five obligations

Our proposal The Continuity Framework v2.0. Every block of it is labeled on the framework page.

01 · Carry
State outlives the session, the tool, the model and the host that created it.
Without it: The project restarts from zero on Monday, or when the model changes.
02 · Govern
Every write names its source and its authority; untrusted input does not become trusted memory.
Without it: A web page an agent read in June is quietly running its decisions in November.
03 · Evolve
A changed fact is superseded with a reason, never silently edited; the record can say what was true when.
Without it: The assistant confidently answers with a fact that stopped being true in March.
04 · Prove
Every carried fact can show where it came from; every change can show who made it, when, and why.
Without it: Nobody can say why the agent believes what it believes, or who changed it.
05 · Forget
Anything carried can be corrected, revoked, expired, deleted and verified gone, including what was derived from it.
Without it: The user deletes a memory and the agent keeps acting on it, because the summary still holds it.

Where to start

If you are new

Twenty minutes, no vocabulary required.

If you build

What to ask of a memory system before you depend on it.

If you research

Where the field stands and what it has not settled.

The field, November 2026

This site first argued in July 2026 that memory was necessary and not sufficient, and that the missing part was governed, time-bounded state with supersession and lineage. Between then and now the research caught up on the core and went past it in three directions this site had not developed: forgetting, identity, and security. The pieces that changed the thesis most:

A 435-paper survey finds the field heavy on storing and retrieving state, light on governing, recovering and relinquishing it.
Always-On Agents: A Survey of Persistent Memory, State, and Governance in LLM Agents · Tianyu Ding and others, 2026
Deleted facts came back in 93% of cases within five ordinary sessions, through checkpoints, artifacts, peer copies and consolidation.
Deleted Is Not Forgotten: Deletion in Persistent LLM Agents Fails by Resurrection, Not Only by Residue · Shulin Lu, 2026
Deleting a memory left its influence in summaries, plans and the cache unchanged; only provenance-guided replay matched a full reset.
Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM Agents · Chao Yao and others, 2026
Soft-deleted vectors in three popular databases were read off disk and inverted back to names, locations and patient data.
Ghost Vectors: Soft-Deleted Embeddings Remain Reconstructible in HNSW Vector Databases · Chandranil Chakraborttii and others, 2026
Memory poisoning traced to the write path: stored entries lose where they came from, so untrusted content is recalled as trusted knowledge.
From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents · Pritam Dash and others, 2026
An agent's identity separated from its model, harness and host, so that swapping any of them is a migration, not a new agent.
Runtime-Independent Persistent Agents: Preserving Identity, Memory, and Code Across Models, Harnesses, and Servers · Zhenyu Zhao and others, 2026
A W3C community group chartered to make agent memory portable and verifiable across vendors; an IETF draft distinguishes cryptographic from access-control erasure.
AI Agent Memory Interoperability Community Group, Charter v1.0 · W3C Community Group (chair: Russell Jackson), 2026

The honest scorecard on this site's own thinking, including where the field was ahead, is at the end of the thesis. The whole set of sources, sixty-odd and growing, is the research library.

How the record uses it

The rest of this site is a directory: assistants, coding agents, memory systems, open models and inference APIs, each graded from the maker's own pages. The framework is the lens those grades are read through. On the memory-systems shelf, export and model independence are the carry obligation, scoping is govern, and the delete test is rung 5 of the erasure ladder, no higher. On the Scorecard, what an assistant keeps, whether you can take it with you, and whether you can delete it are the same questions asked of consumer products. Each graded page says which obligations its tests reach and which they do not.

The directory shows what exists. This section says what good architecture looks like and what is still missing. Keeping those apart is the point: a product can score well on every test here and still not be a continuity layer, and this site says so.

Questions

What is a continuity layer in AI?: A continuity layer is the part of an AI system that carries forward governed state across sessions, tools, models, and hosts: what is currently true, what changed, why it changed, by whose authority, and which evidence supports it. It sits above memory and retrieval, and below agent execution, turning stored history into usable, auditable continuity; and it can correct, supersede, revoke, expire, and forget what it carries, with evidence that it did.
Is a continuity layer the same as AI memory?: No. Memory stores the past; a continuity layer governs what of it is still true, who may change it, why it changed, and whether it should be carried forward at all, and it can forget. Memory is necessary for continuity and not sufficient. Products like Mem0, Zep, Letta and the assistants' own memory features are memory systems; this site grades them on the questions a continuity layer would ask.
Why is forgetting part of continuity?: Because a record that cannot be corrected, revoked, expired, deleted and verified gone is a liability, not a working relationship. The 2026 research shows that deleting a stored memory leaves its influence in summaries, plans and indexes, and that deleted facts return within a few sessions through checkpoints, copies and consolidation. The erasure ladder names twelve operations so a “delete” claim can be read for what it guarantees.
Is the Continuity Framework an industry standard?: No. It is this site's proposal, versioned and signed, and every block of it is labeled as established, emerging, ours or a hypothesis. The established vocabulary it rests on comes from the sources in the research library: W3C PROV for provenance, SQL:2011 for bi-temporal data, NIST SP 800-88 for cryptographic erase, and the 2026 surveys for lifecycle and governance.

The Continuity Layer

The directory shows what exists. This section explains why it matters and what good architecture looks like. The two meet on the graded shelves, where every memory system and assistant is read through the same five obligations.