Big data analytics in datacenter platforms and data intensive simulations in exascale computing environments create the need for massive main memory capacities, on the order of terabytes, to boost application performance. To satisfy these requirements, memory hierarchies become more complex, incorporating emerging types of technologies or disaggregation techniques to offset the skyrocketing cost that DRAM-only systems would impose. As we shift away from traditional memory hierarchies, the effectiveness of existing data management solutions decreases, as these have not provisioned against the even bigger disparity in the access speeds of the heterogeneous components that are now part of the memory subsystem. Additionally, system-level configuration knobs need to be re-tuned to adjust to the speeds of the newly introduced memory hardware. In the face of this complexity, conventional approaches to designing data management solutions with empirically-derived configuration parameters become impractical. This makes the case for leveraging machine intelligence in building a new generation of data management solutions for hybrid memory systems. This thesis identifies the machine intelligent methods that can be effective for and practically integrated with system-level memory management, and demonstrates their importance through the design of new components of the memory management stack; from system-level support for configuring stack parameters to memory scheduling.