• memory augmented neural networks
• neural networks
• reinforcement learning

This work explores the capabilities of the current Reinforcement Learning algorithms and the Memory Augmented Architectures in tackling tasks that requires the use of external memory. It also proposes a new architecture that improves the model capabilities in solving this type of tasks. The models have been tested on two different grid-based maze tasks, where the agent needs to remember some information to act accordingly later. We compare our proposed model with Neural Turing Machines and Long Short Term Memory in their capabilities of solving this task.