In the field of robotics, the ability to perform long-term tasks within human-centric environments is essential for applications such as elder care and logistics management. Current methodologies, such as diffusion-based planners and reinforcement learning techniques, often exhibit limitations in adaptability and managing long-horizon tasks. To mitigate these challenges, we introduce MambaSkill, a framework designed to enhance the generation of skill sequences by modeling long-term dependencies through the implementation of Mamba. The principal innovations of MambaSkill comprise a Mamba-based autoencoder for skill abstraction and a dual-level Mamba-enhanced autoregressive generator for the prediction of dynamic skill sequences. These advancements allow MambaSkill to generate motion primitives of variable lengths while maintaining coherence across extended tasks. Empirical results indicate that MambaSkill outperforms prior methodologies, achieving a 7.5% improvement in long-horizon tasks and a 1.2% gain in ordinary tasks, along with a reduction in GPU resource consumption. These results validate its effectiveness in long-term robotic manipulation and general task performance.