We propose LEBA, a framework that mitigates forgetting and enhances generalization in multi-domain continual learning via Schrödinger bridge-based optimization and progressive adapter ensembling.