Inherited, Not Remembered: Lifecycle Consolidation for Successor Language Agents

Long-lived language agents increasingly accumulate experience during deployment, yet current model lifecycles provide no principled account of how that experience should be handled when one model replaces another. Continuous updating risks drift and contamination; external memory preserves traces without necessarily producing capabilities; and behavioral transfer does not determine what should be inherited or forgotten. This paper proposes lifecycle consolidation: an explicit phase in which predecessor experience is audited, abstracted into generalizable faculties (experience-derived capacities), and transferred to a successor while episode-specific traces are discarded or archived. We define the episode/faculty distinction, present a six-stage reference architecture, and derive falsifiable comparisons among successors trained from scratch, continuously fine-tuned models, episode-transfer successors, and lifecycle-consolidated successors. The central prediction is a dissociation: a successor retains useful capacities derived from deployment history without recalling their source episodes. Model replacement thereby becomes a problem of selective inheritance rather than simple substitution.