In this paper we show that, thanks to the features of Continuous-Time Recurrent Neural Network (CTRNNs), whose neurons act at different activation level of time, the controller of a bot is capable of learning during runtime in the UT2004 videogame environment without changing any of the network parameters. This behaviour has been described in simple forms of life (e.g. examples of the small nematode worm C. elegans, evidence for the formation of associations between temperatures and food has been known for quite some time) and, in this paper, we adapt these ideas applying them to a current commercial video game. The particular experimental conditions are as following: A bot is evolved in a task where it needs to search and discriminate its base camp and the enemy’s camp and associate them with the altitude where the camp is, depending on its experience. The task requires either instrumental or classical conditioned response to be learned. In the last part of the paper is analytically analyzed the best-evolved agent’s behaviour and, also, it is explained how some ideas and methodologies that come from models of biological theory, can be easily adapted to the control of virtual characters or “bots”, i.e. synthetic agents with human behaviour.