Continuity Layers for Long-Running LLM Assistants: Deterministic Checkpointing, Rehydration, and Fail-Closed Gates

This preprint introduces a continuity-layer architecture for long-running LLM assistants. We argue that many apparent “memory” failures are continuity failures caused by lossy context mutation (summarization/truncation), missing state lineage, and non-deterministic recovery behavior. The paper presents a deterministic control-plane pattern using checkpointing, artifact-grounded rehydration, integrity/lineage/freshness validation, and fail-closed gating for state-critical transitions. We provide a concrete checkpoint schema, PASS/FAIL payload examples, and an evaluation protocol with baselines and metrics (rehydrate success, determinism rate, fail-closed coverage, false fail-closed rate, and recovery MTTR). Current evidence is scoped to literature-backed motivation and internal qualitative observations; quantitative benchmark publication is proposed as future work. Primary contribution: reframing “LLM memory” as an engineered continuity systems problem with auditable, testable controls.