The goal of this PhD is to explore a minimal model of decision making using a simulated agent in a contiguous environment (T-Maze like). The goal for the agent is to learn to alternate between left and right, independently of the geometry of the maze, even though topology remains the same. This will be done using an echo state network of limited size in order to be able to perform a thorough analysis of its dynamics and representations from three different perspectives (sensory-motor space, external behavior and neural activity). The goal is to find the conditions for the emergence of concepts such as left and right using a manifold-based approach and to prove for their existence independently an external observer.