Memory-Free Continual Learning with Null Space Adaptation for Zero-Shot Vision-Language Models