AI-assisted research systems increasingly preserve more than isolated prompts and final outputs. They maintain memory layers, record action traces, package workflow provenance, generate candidate artifacts, and in some cases produce full paper-like objects. Karpathy-style AutoResearch and LLM Wiki patterns make this shift visible: one foregrounds replayable experiment loops, while the other foregrounds persistent managed memory between raw sources and later query. This paper argues that replay and memory are valuable but insufficient for research continuation unless they are composed into governed research state. Replayability means that a record is sufficient to reconstruct or rerun a sequence of actions. Resumability means that a record is sufficient to restore operative research state for accountable continuation. The distinction is elementary but methodologically underused in discussions of AI-assisted research infrastructure. Existing literatures on autonomous research agents, provenance, research objects, workflow systems, process mining, scholarly knowledge graphs, and agent memory each address important layers of this problem. The Reflexive Laboratory is used as a worked comparator, not a universal implementation. Its prior models of transcript-to-state derivation, bounded autoresearch, execution versus state sufficiency, artifact graphs, canonicality, and sanity-check operators specify a composed layer of evidence-mediated working state, artifact identity, authority, validation, human admissibility judgment, and canonical publication status. The contribution is a research-continuation layer model, not a replacement for existing infrastructure standards. This release package includes the manuscript, source files, figures, tables, bibliography, provenance notes, AI-use note, and a transcript supplement documenting part of the conceptual-development process.
Peter Bell (Fri,) studied this question.