This record describes an ongoing research program proposing a continual-learning agent architecture inspired by biological separation between rapid online adaptation and offline consolidation. The method alternates a Day phase (stable inference with bounded, parameter-efficient ephemeral adaptation and episodic logging) and a Night phase (compute-accounted offline consolidation from an episodic buffer prioritized by a scalar valence signal). To prevent category errors and “symbol hijacking,” the paper includes an interaction-first formal core introducing an interaction-composition operator distinct from arithmetic, and a strict lane-separation / claim-ladder that keeps speculative ideas quarantined from testable engineering claims. The work preregisters a falsifiability protocol: compute matching, baselines, mandatory ablations, and explicit failure conditions. A parallel prototype track targets edge feasibility under thermal, memory, and latency constraints, with planned reporting on token latency distributions, throttling/thermal traces, memory/KV-cache behavior, IO throughput, and Night duty-cycle gating. Empirical benchmark results and open implementation artifacts are in preparation and will be released as the prototype stabilizes.