Persistent memory through triple-loop consolidation under stochastic unit turnover

Dissipative cognitive architectures maintain computation through continuous energy expenditure, where units that exhaust their energy are stochastically replaced with fresh random state. This creates a fundamental challenge: how can persistent, context-specific memory survive when all learnable state is periodically destroyed? Existing memory mechanisms—including elastic weight consolidation, synaptic intelligence, and surprise-driven gating—rely on gradient computation and are inapplicable to dissipative systems that do not perform it. We introduce Deep Memory (DM), a backpropagation-free persistent memory mechanism operating through a triple-loop consolidation cycle: (1) recording of expert-specific content centroids during active computation, (2) seeding of replaced units with stored representations, and (3) stabilization through continuous re-entry that counteracts dissipative drift. We demonstrate that discrete expert routing via Mixture-of-Experts (MoE) gating is required, in the regimes tested, for DM functionality, preventing the centroid convergence that would render stored memories identical. Across 972 simulation runs spanning thirteen experimental blocks, we establish: (i) removing stable context–expert binding removes structural specialization (MI = 1.10 vs. 0.001 under binding destruction; n = 91); (ii) DM achieves persistent representation quality R = 0.984 compared to 0.385 without memory ( n = 16); (iii) continuous seeding reconstructs representations after interference ( R recon = 0.978; one-shot seeding fails; n = 30); (iv) the mechanism operates within a well-characterized ( K , ρ) envelope with identifiable degradation boundaries ( n = 350); (v) a factorial ablation isolates recording × seeding as the minimal critical dyad ( n = 40); and (vi) six confirmatory experiments—including single-factor causal ablation, mediation analysis, and comparison with associative and reservoir baselines (Hopfield, ESN) under matched turnover—validate what the mechanism requires and how it compares with those baselines ( n = 370). Results (ii)–(v) rest on pilot-scale seed counts (4–5 seeds per condition); at these counts no paired non-parametric test can reach p < 0.05 whatever the effect size, so we report them as estimates of effect size and direction rather than as confirmatory tests. Within these limits, DM is a falsifiable, bounded mechanism for persistent memory in backpropagation-free cognitive systems. Its relation to hippocampal consolidation is a functional analogy only: DM shares the premise of a store held apart from the computing substrate, but assigns the fast and slow roles in the opposite direction to complementary learning systems theory.