Abstract
Memories for temporally extended sequences can be used adaptively to predict future events on multiple timescales, a function that relies on the hippocampus. For such predictions to be useful, they should be updated when environments change. We investigated how and when new learning shapes hippocampal representations of temporally extended sequences, and how this updating relates to flexible predictions about future events. Human participants learned sequences of environments in immersive virtual reality. They then learned novel environment transitions connecting previously separate sequences. During subsequent fMRI, participants predicted multiple steps into the future in both the newly connected sequence and control sequences that remained separate. The hippocampus integrated representations of the connected sequence, such that activity patterns became more similar across trials for the connected sequence vs. the unconnected sequences. These integrated sequence representations in the hippocampus emerged soon after learning, incorporated representations of the initial sequences as well as new activity patterns not previously present in either sequence, and predicted participants' ability to update their predictions in behavior. Together, these results advance our understanding of how structured knowledge dynamically emerges in service of adaptive behavior.