This paper explores the integration of the Mamba framework into the Decision Transformer (DT) architecture to enhance its performance in sequential decision-making tasks. The Mamba framework, known for its efficient and effective sequence modeling capabilities, introduces a selective structured state space model that can selectively extract essential information while filtering out irrelevant noise. The authors introduce *Decision Mamba* by substituting the causal self-attention module in DT with Mamba, aiming to improve the model's ability to capture complex dependencies and nuances in sequential decision-making tasks. The study systematically evaluates the integration of Mamba into DT through a series of experiments across various decision-making environments, including continuous control tasks from the D4RL benchmark and discrete Atari games. The results show that Decision Mamba performs competitively with existing DT-type models, suggesting the effectiveness of Mamba in RL tasks. The paper also includes ablation studies to investigate the contributions of the channel-mixing layers and the context length in the Mamba block. The authors conclude that while the integration of Mamba into DT does not significantly enhance efficiency due to the hardware interactions involved in RL tasks, it offers valuable insights into the interplay between different architectural components and their impact on model performance in complex decision-making environments. Future work could explore further improvements in parameter tuning and adapting the network architecture to better suit the RL data structure.This paper explores the integration of the Mamba framework into the Decision Transformer (DT) architecture to enhance its performance in sequential decision-making tasks. The Mamba framework, known for its efficient and effective sequence modeling capabilities, introduces a selective structured state space model that can selectively extract essential information while filtering out irrelevant noise. The authors introduce *Decision Mamba* by substituting the causal self-attention module in DT with Mamba, aiming to improve the model's ability to capture complex dependencies and nuances in sequential decision-making tasks. The study systematically evaluates the integration of Mamba into DT through a series of experiments across various decision-making environments, including continuous control tasks from the D4RL benchmark and discrete Atari games. The results show that Decision Mamba performs competitively with existing DT-type models, suggesting the effectiveness of Mamba in RL tasks. The paper also includes ablation studies to investigate the contributions of the channel-mixing layers and the context length in the Mamba block. The authors conclude that while the integration of Mamba into DT does not significantly enhance efficiency due to the hardware interactions involved in RL tasks, it offers valuable insights into the interplay between different architectural components and their impact on model performance in complex decision-making environments. Future work could explore further improvements in parameter tuning and adapting the network architecture to better suit the RL data structure.