The hippocampus is thought to organise knowledge about the world by binding stimuli into representations of discrete states, which can be utilised to guide flexible behaviour. In the laboratory, such states are often characterised as distinct spatial locations and contexts. However, real-life experiences are often defined by latent non-spatial contexts. Despite a hypothesised key role, the extent to which the hippocampus may form representations of such abstract states is unclear. To study the role of hippocampus in such latent state representations, we used a 2-armed bandit task in mice. In this task, two abstract states are defined by blocks of different reward contingencies associated with two actions, that unsignalled to the mouse reverse throughout the course of the task. Thus, due to probabilistic rewards and unsignalled block reversals, mice must integrate evidence over several trials to infer the current latent state to guide their behaviour. Using computational modelling and recordings of midbrain dopamine release, we first show that mice perform inference over two latent states in this task. We next looked for the representation of these abstract states in ventral hippocampus (vH) using Miniscope imaging of single neuron activity in freely moving mice. We found that vH encodes these states in a manner similar to different spatial contexts, maintaining differentiable representations of each latent state, despite the spatial location remaining unchanged. Finally, using chronic ablation of pyramidal neurons in vH, we show that hippocampal activity is necessary for the use of latent state inference during performance of the task.