As the memory capacity demands of memory-intensive applications continue to grow, tiered memory architectures that incorporate larger-capacity, lower-cost slow memory layers have become an important approach for reducing data center hardware costs. However, due to the significantly higher access latency and lower bandwidth of slow memory, memory management policies play a critical role in overall system performance. During execution, memory-intensive applications exhibit highly dynamic memory access behavior, while most existing tiered memory management schemes rely on static migration policies that are difficult to adapt to phase-varying access patterns. Such static strategies often lead to ineffective or excessive page migrations, thus offsetting the potential performance benefits of tiered memory architectures. To address these challenges, this study proposes a hot-page identification mechanism that integrates access frequency and access recency to more accurately characterize the comprehensive access value of memory pages. Building upon this mechanism, the study further designs a reinforcement learning-based dynamic tiered memory management method, which adaptively regulates cross-tier page migration intensity through policy learning at runtime. The proposed method is implemented and evaluated on a real DRAM-PMEM tiered memory platform. Experimental results demonstrate that the proposed method significantly reduces ineffective page migration overhead while consistently outperforming existing tiered memory management solutions across a variety of representative memory-intensive workloads.