Efficient deep neural networks are essential for deployment on resource-constrained platforms. In this paper, we propose an improved structured pruning framework that jointly considers L1-norm sparsity and peak memory usage to effectively reduce both computational cost and runtime memory overhead. Our method introduces a memory-aware criterion that iteratively selects channel configurations by minimizing a combined loss of L1-norm and peak memory consumption. We validate our approach on benchmark datasets such as CIFAR-10 using VGG16 with Batch Normalization (VGG16-BN), ResNet, and Transformer architectures. To ensure practical applicability, we implement true channel removal along with synchronized BatchNorm alignment and dynamic reconstruction of the inference graph. Experimental results demonstrate that the proposed method significantly reduces peak memory usage while maintaining high accuracy, making it suitable for deployment on edge devices. Additionally, the framework supports an optional memory loss toggle, enabling flexible analysis of the trade-off between memory consumption and model accuracy.