Motor condition-monitoring data accumulate continuously during operation, while fault categories evolve from unobserved to partially confirmed and progressively incorporated into the diagnostic space. Existing long-tailed fault diagnosis methods are mostly based on static datasets and cannot capture continuous sample accumulation, gradual category confirmation, and changing diagnostic task boundaries. This paper proposes a dynamic long-tailed motor fault diagnosis method for continuous sample accumulation. First, a three-stage framework comprising anomaly detection, open-set long-tailed diagnosis, and closed-set long-tailed diagnosis is constructed to establish normal-state boundaries, recognize known faults while rejecting unknown faults, and perform closed-set long-tailed recognition across all categories, respectively. Second, a complementary strategy of generative augmentation and discriminative optimization generates pseudo-samples for known tail classes and adaptively regulates class-wise pseudo-sample injection based on class compensation demand, diagnostic difficulty, and the discriminative confidence of reliable pseudo-samples. A weighted training constraint further uses generated samples to assist in optimizing decision boundaries for tail classes. Experimental results show that the proposed method adapts to dynamic long-tailed diagnosis under continuous sample accumulation while balancing open-set unknown-fault rejection and closed-set tail-fault recognition.