The theory of Bayesian inference presents a normative approach to understanding how animals and humans learn about their environment. To demonstrate this, I will start by introducing the theory and show as an example how it explains aspects of human chunk learning in a visual learning paradigm that cannot be captured by traditional associative learning accounts. I will then turn to a complementary view of Bayesian inference: how it can be used as a data analysis tool to estimate mental representations of object classes from simple binary response data collected in psychophysical experiments. Such methods can be used to track as humans develop complex internal representations, with minimal changes to already existing experimental paradigms. Finally, I will take a step back, and place learning and memory within the wider context of behavioural economics. I will argue that even though Bayesian inference offers a statistically optimal way for learning, the representations it learns — internal models — can be highly inefficient for decision making. This leaves room for qualitatively different ways of learning to be advantageous under some ecologically relevant conditions. I will show how one such alternative, episodic memory, can be understood as a better way to support optimal decision making under risk and uncertainty in complex environments, and how this normative view of episodic memory accounts for many of its behavioural and neural correlates. These studies together provide a principled framework to explore complex learning and developmental phenomena reported in humans and animals.