Continual learning in skeleton-based action recognition poses a fundamental challenge because of the need for models to incorporate additional classes without catastrophic forgetting of previously learned actions. One effective strategy to mitigate forgetting is the use of replay samples that we draw from past experiences. However, the optimal balance between newly introduced data and replayed samples—termed the replay ratio—remains unclear. In this paper, we systematically investigate the impact of varying replay ratios on performance and training efficiency by using a spatial-temporal graph convolutional network (ST-GCN) model that we train on the NTU-RGB + D skeleton dataset. We evaluate a wide range of ratios and compare recognition performance and total training time to identify the most favorable trade-off. Our results show that while extremely low replay ratios suffer from significant performance degradation, higher ratios incur substantial training overhead. Particularly, 10–30% replay ratios appear to provide a favorable balance, maintaining nearly optimal recognition performance while considerably reducing training time compared to higher replay ratios. These findings provide actionable guidance for designing more resource-efficient continual learning pipelines in skeleton-based action recognition.