Humans and animals learn from reward but they also learn by observing statistical relationships in the world. It is the coalescence of these learning mechanisms that shapes our ability to produce complex goal-directed behaviours. While much is known about the neural encoding of updating signals during learning, there is relatively little knowledge on where and how learnt representations are stored. The first study I will present explores the neural representations or ‘associative structures’ created by multiple different learning mechanisms using human fMRI. We find that knowledge encoded via model-free RL is dissociable, neurally, from the encoding of statistically learnt relationships. One advantage of acquiring relational knowledge is that it allows us to behave adaptively in new situations and make inferences about never previously experienced options. In the second study I will examine whether macaque monkeys can make inferences about novel choice options and show that they recruit a hexagonal map-like coding scheme to represent relationships in an abstract option space. The third study will take a different approach and focus on an anatomical circuit centred on the amygdala in a large cohort of healthy human participants (Human Connectome Project) to examine whether measures of functional coupling of specific amygdala nuclei can predict markers of mental well-being.