Connecting computational neuroscience with mechanistic interpretability

> This study investigates whether the Replay-Gated Cascade Consolidation (RGCC) mechanism extends beyond spiking neural networks to transformer-based language models. We test the hypothesis that sequential learning in transformers exhibits the same replay-driven consolidation dynamics predicted by RGCC, including competition for shared parameters and an encoding-order consolidation gradient. By analyzing sequential fine-tuning and replay strategies, we evaluate whether RGCC provides a unified mechanistic explanation for memory consolidation and catastrophic forgetting across fundamentally different neural architectures.