This paper introduces Decision Mamba, a novel architecture that integrates the Mamba framework into the Decision Transformer (DT) for improved sequential decision-making in reinforcement learning (RL). Decision Transformer uses causal self-attention to model sequences of states, actions, and rewards, but this paper explores the integration of Mamba, a sequence modeling framework with efficient and effective state space modeling capabilities. The Mamba framework introduces a data-dependent selection mechanism and hardware-aware design to address data- and time-invariant issues in prior state space models, enabling selective information extraction and superior sequential modeling performance. The paper evaluates the integration of Mamba into DT through experiments across various decision-making environments, comparing the modified Decision Transformer (Decision Mamba) with its traditional counterpart. The results show that Decision Mamba is competitive with existing DT-type models, suggesting that Mamba can enhance RL performance. The architecture of Decision Mamba replaces the self-attention module in DT with the Mamba block, which is designed to capture complex dependencies and nuances in sequential decision-making tasks. The paper also discusses the training and inference processes of Decision Mamba, showing that it can effectively model sequences of states, actions, and rewards. The results from experiments on continuous OpenAI Gym and discrete Atari environments demonstrate that Decision Mamba performs well, with comparable or improved performance compared to other models like Decision S4 and Decision ConvFormer. Ablation studies show that the Mamba block is sufficient for RL sequence modeling tasks, and that the context length K significantly affects performance in different environments. The paper concludes that Decision Mamba offers a promising approach for improving the performance of Transformer-based models in RL scenarios. However, it also highlights the need for further research on the efficiency of the model and the potential for adapting the network architecture to better suit RL data structures. The study contributes to the advancement of sequential decision-making models, suggesting that the architecture and training methodology of neural networks can significantly impact their performance in complex tasks.This paper introduces Decision Mamba, a novel architecture that integrates the Mamba framework into the Decision Transformer (DT) for improved sequential decision-making in reinforcement learning (RL). Decision Transformer uses causal self-attention to model sequences of states, actions, and rewards, but this paper explores the integration of Mamba, a sequence modeling framework with efficient and effective state space modeling capabilities. The Mamba framework introduces a data-dependent selection mechanism and hardware-aware design to address data- and time-invariant issues in prior state space models, enabling selective information extraction and superior sequential modeling performance. The paper evaluates the integration of Mamba into DT through experiments across various decision-making environments, comparing the modified Decision Transformer (Decision Mamba) with its traditional counterpart. The results show that Decision Mamba is competitive with existing DT-type models, suggesting that Mamba can enhance RL performance. The architecture of Decision Mamba replaces the self-attention module in DT with the Mamba block, which is designed to capture complex dependencies and nuances in sequential decision-making tasks. The paper also discusses the training and inference processes of Decision Mamba, showing that it can effectively model sequences of states, actions, and rewards. The results from experiments on continuous OpenAI Gym and discrete Atari environments demonstrate that Decision Mamba performs well, with comparable or improved performance compared to other models like Decision S4 and Decision ConvFormer. Ablation studies show that the Mamba block is sufficient for RL sequence modeling tasks, and that the context length K significantly affects performance in different environments. The paper concludes that Decision Mamba offers a promising approach for improving the performance of Transformer-based models in RL scenarios. However, it also highlights the need for further research on the efficiency of the model and the potential for adapting the network architecture to better suit RL data structures. The study contributes to the advancement of sequential decision-making models, suggesting that the architecture and training methodology of neural networks can significantly impact their performance in complex tasks.