Deep meta-reinforcement learning (meta-RL) is introduced as a novel approach to enable deep reinforcement learning (RL) systems to adapt rapidly to new tasks. The method involves training a recurrent neural network (RNN) using one RL algorithm, while the RNN's dynamics implement a second, separate RL procedure. This second procedure can differ significantly from the original algorithm and is configured to exploit structure in the training domain. The approach is validated through seven proof-of-concept experiments, each examining a key aspect of deep meta-RL. The experiments demonstrate that the learned RL procedure can adapt to new tasks more efficiently than traditional RL methods, and that it can generalize across different task distributions. The framework is also discussed in terms of its potential implications for neuroscience. The experiments include studies on bandit problems and Markov decision problems, showing that meta-RL can learn adaptive strategies for exploration and exploitation, and can generalize to new tasks with different structures. The results suggest that the learned RL procedure can differ significantly from the original algorithm used to train the network weights, and that it can exploit consistent task structure to improve performance. The approach is also shown to be scalable to complex navigation tasks with rich visual inputs.Deep meta-reinforcement learning (meta-RL) is introduced as a novel approach to enable deep reinforcement learning (RL) systems to adapt rapidly to new tasks. The method involves training a recurrent neural network (RNN) using one RL algorithm, while the RNN's dynamics implement a second, separate RL procedure. This second procedure can differ significantly from the original algorithm and is configured to exploit structure in the training domain. The approach is validated through seven proof-of-concept experiments, each examining a key aspect of deep meta-RL. The experiments demonstrate that the learned RL procedure can adapt to new tasks more efficiently than traditional RL methods, and that it can generalize across different task distributions. The framework is also discussed in terms of its potential implications for neuroscience. The experiments include studies on bandit problems and Markov decision problems, showing that meta-RL can learn adaptive strategies for exploration and exploitation, and can generalize to new tasks with different structures. The results suggest that the learned RL procedure can differ significantly from the original algorithm used to train the network weights, and that it can exploit consistent task structure to improve performance. The approach is also shown to be scalable to complex navigation tasks with rich visual inputs.