As the KEM algorithm primarily recommended by the National Institute of Standards and Technology (NIST), ML-KEM (the standard derived from the CRYSTALS-Kyber) can provide post-quantum security for embedded devices such as ARM Cortex-M4 devices, which are widely deployed for Internet of Things (IoT) applications. However, the considerable memory footprint presents a significant bottleneck for deployment on resource-constrained devices. To address this issue, we optimize the memory usage of ML-KEM implementation on embedded devices. By combining the optimization techniques of segmented loading, on-the-fly generation, buffer reuse, and streaming output, we achieve a memory-efficient implementation of ML-KEM that substantially reduces the peak memory usage (memory footprint) during encapsulation and decapsulation. The evaluation results show that the RAM usage is reduced by up to 25.29% for encapsulation and up to 37.61% for decapsulation compared to the stack-optimized implementation in PQM4 library, with at most about 0.11% latency overhead. This work demonstrates the feasibility of deploying post-quantum (lattice-based) cryptographic schemes on resource-constrained embedded platforms and provides a practical foundation for their adoption in real-world IoT devices.