AI chips face on-chip memory limits in deep learning. Current optimization methods focus on static computation graphs, leaving room to improve memory efficiency for dynamic graphs. To overcome this limitation, a memory optimization framework for control-flow computation graphs was developed. The framework realized operator-level memory reuse within subgraphs and further achieved recursive reuse across subgraphs by exploiting control-flow characteristics. In addition, a ping-pong buffering strategy for weight data was introduced to mitigate the memory wall between on-chip and off-chip memory, thereby allowing overlapping of memory access and computation operations within subgraphs. Validation on the domestic LUNA AI chip has demonstrated that the proposed framework improves on-chip memory utilization by 5.9% compared with existing methods. Moreover, the strategy effectively alleviates the memory wall problem by reducing data transfer time between on-chip and off-chip memory, resulting in execution efficiency improvements of up to 29%.