When Experience Leaves a Trace: Consolidation-Dependent Persistence in Artificial Agents

Abstract

Across minimal neural networks and small transformer models, we demonstrate that experience ordering alone can produce durable, irreversible behavioral divergence in artificial agents--but only when learning is consolidated into internal parameters rather than externally scaffolded. We test six architectural variants against four operational diagnostics: deletion resistance, path dependence, irreversibility, and preference stability. Systems relying on external memory (context windows, retrieval stores) fail all tests-- deleting external records eliminates apparent persistence entirely. Systems with surprise-gated plasticity pass deletion and path-dependence tests, with replay-based consolidation amplifying behavioral divergence 8.6-fold between differently trained agents. Only architectures incorporating explicit viability variables pass preference stability, consistently sacrificing external reward to preserve internal states. These results localize a boundary gap: current architectures preserve designer-specified viability variables but do not discover which internal states matter for their own persistence. The diagnostics provide operational tools for distinguishing adaptive systems with endogenous memory from externally scaffolded tools, and identify concrete architectural requirements for crossing this boundary.