Research on architectures for efficient long-sequence processing, primarily state-space models and their hybrids with transformers. We investigate how the design of dynamic memory affects information retention and updating, handling of dependencies across different time scales, and training stability, as well as which operations can be performed via a compact state and which require direct access to the full context. On this basis, we develop and analyze novel approaches to layer-wise state management, memory mechanisms, and hybrid architectures that combine the strengths of recurrent processing and self-attention. The goal of this research direction is to formulate principles for building sequence models capable of effectively handling long contexts at lower computational cost without significant loss of quality or model capability.