Session State
Continuity
& Anchor Replay Family.

Eight sections — from the copy-paste continuity failure to the buildable Family Pack that preserves epistemic state across session and model boundaries.

Take any long-running piece of work you have done with an AI system over the last several months. A specification you developed across twenty sessions. A research argument you built up over a quarter. An analysis that matured incrementally, each session resuming where the last one left off. Now imagine moving the entire history into a different model — copying the conversation, pasting it in, and asking the new model to continue.

The new model can read every word. It can summarize the topic. It can answer questions about what was discussed. And yet anyone who has actually attempted this knows: the work does not continue. It restarts on the same topic.

What is lost in the copy is not content. The content transferred. What is lost is everything about the content that gave the work its shape:

• Which claims were established and which were still open. The text reads the same either way; the distinction is gone.
• Which questions were intentionally left unresolved versus which simply never came up. The new session cannot tell them apart.
• Which work was explicitly designated to continue. The obligation does not survive as an obligation; it survives, if at all, as a sentence among other sentences.
• Where each claim came from — retrieved from a document, derived by the prior model, declared by you as a working assumption. After the copy, all of it reads with the same weight.
• Which conclusions had been reached collaboratively, and which were proposals that you had not yet endorsed.

This is what we mean by epistemic continuity: not the content of the work, but the structure of the thinking — what is settled, what is open, what is provisional, what is owed, what came from where. Inside a single session, this structure exists implicitly in the working memory of the conversation. Across a session boundary, with current tools, none of it crosses.

In long-running serious work, the difference is not cosmetic. A safety analysis where intentional openness silently became settled is no longer the same analysis. A specification where a designated continuation obligation quietly disappeared is missing work it was supposed to contain. A research argument where the provenance of every claim has been flattened cannot be reviewed for what it actually rests on. The copy carried the words. It did not carry the work.

The cost of broken epistemic continuity is delayed and compounding. Nothing crashes at the boundary. The next session reads the carried content fluently. The work appears to continue. The damage is in what the next session no longer knows about what the prior session was doing.

In each case, no one acted in bad faith. The prior work was real. The next session was real. What was missing was not vigilance — it was the machinery that lets the thinking, not just the text, cross the boundary.

Three approaches dominate how long-running AI work bridges sessions today. All three move content across the boundary. None of them moves the structure of the thinking.

Conversation history and scrollback Within a session, you can re-read prior turns. This preserves visible content for the current session. It does not span to the next session, and even within the session it does not make explicit what is settled, what is open, what is provisional, or where each claim originated. It is content, not epistemic state.

Provider-side memory features Some platforms persist user context across sessions on the provider’s infrastructure. These features improve usability. They carry text and references; they do not carry the structure of the prior session’s thinking. Whether a question was deliberately open or accidentally unanswered, whether a conclusion was reached or merely proposed — distinctions like these are not what these features were built to preserve.

Manual handover documents The most common approach for serious work is a handover document — a careful summary you maintain between sessions and paste into the next one. The document carries content. It does not carry the structure underneath the content. There is no place in a flat summary for “this is still open by design,” “this is a working assumption I have not endorsed,” “this is an obligation to continue.” These distinctions either survive accidentally, in how a sentence happens to be phrased, or they do not survive.

The question current approaches do not answer is not “how do I carry the content forward.” Current approaches answer that question adequately. The question they do not answer is: How does the structure of the prior session’s thinking — what is settled, what is open, what is owed, where each claim came from — continue into the next session? This is the gap SSCAR addresses.

SSCAR — the Session State Continuity & Anchor Replay Family — is a specification for governing the epistemic continuity of long-running AI work across session boundaries, including across changes of model. The family does not produce content, generate summaries, or decide what the prior work should have said. It governs whether the next session inherits the structure of the prior session’s thinking — not just its words.

The family preserves the continuity of the work. It does not judge whether the work was correct, repair anchors that fail to support continuity, or generate substitute structure for what the prior session did not record. Where continuity cannot be established, the family causes the next session to fail closed rather than proceed on a foundation whose shape has been lost.

The specification is ready-to-build and execution-neutral. It does not prescribe a programming language, framework, or runtime.

The family provides four guarantees. These are structural properties of the specification.

How these guarantees are achieved is the content of the specification.

The family applies wherever AI work persists across sessions and the integrity of that persistence — not just its content — matters.

The §07 content reduces to two configurations: Component Agent and Family Pack.

Family Pack delivers epistemic continuity across session and model boundaries

The Session State Continuity & Anchor Replay Family Pack is a complete set of ready-to-build, execution-neutral normative specifications and operational connector annexes. It is not a concept paper, not a framework, not a set of guidelines, and not an architecture diagram.

It is the document set you hand to an engineering team when you want them to build the family of components that produces the guarantees described in this whitepaper. The engineering team does not need to make design decisions about what each agent should do, what the integrator should compose, how the connector wiring should work, or how the family should fail. Those decisions are made. The specifications and annexes contain them.

The Session State Continuity & Anchor Replay Family specifications include everything necessary for implementation: each component agent's identity and boundaries, contracts with external systems, complete internal logic, input rules, evidence obligations, failure handling, resumption conditions, and acceptance criteria — plus the integrator's family-level decision logic, the producer-consumer wiring, and the runnable family connector test suite.

Nothing is left undefined. Nothing is left to interpretation. Nothing requires your team to reverse-engineer intent from ambiguous prose.

Every whitepaper in the library is free to read. Every specification is distributed free of charge, as a Single-agent Pack or a Family Pack. Free to use, free to verify, free to publish about — free to implement and to embed in your own products, including commercially. The license sets out what is reserved. The whitepaper shows you the problem and the promise. The specifications and annexes give you the solution.