Feedback-regulated dual-role memory consolidation for continual learning: a stability-plasticity framework inspired...

Artificial agents trained on non-stationary task streams often acquire new categories at the cost of older representations. Systems-level accounts of hippocampo-cortical consolidation suggest that stored traces can support later learning in different ways and that consolidation can depend on both new learning and the vulnerability of established knowledge. Motivated by these functional principles, we propose dual-role memory consolidation learning (DRMCL) for class-incremental learning with a bounded replay buffer. DRMCL stores prototype-preserving core samples and decision-sensitive boundary samples, then combines both roles with replay, logit distillation, and task-level feedback control. On Split CIFAR-10 with a pretrained ResNet18 backbone, the clearest benefit occurs when memory is scarce. With 1000 stored samples, average accuracy increases from 0.4823 for Dark Experience Replay++ (DER++) to 0.6905, while forgetting decreases from 0.5734 to 0.2650. At medium and high memory, DRMCL usually lowers forgetting, although DER++ can achieve higher raw accuracy. Ablations show that feedback-regulated replay and distillation account for most of the retention effect, while the core/boundary organization makes the selected memory easier to inspect. An electroencephalography (EEG) brain-computer interface (BCI) feasibility study on BCI Competition IV 2a provides a cross-domain check. In a five-seed subject-1 experiment, DRMCL remains close to ER and DER++ rather than separating from them, and session spectral/coherence shifts vary across subjects. The results support DRMCL as a stability-oriented computational framework; they do not establish a circuit-level model of hippocampo-cortical consolidation or a new EEG decoding benchmark.