We investigated how humans encode large-scale spatial environments using a virtual taxi game. We hypothesized that if 2 connected neighborhoods are explored jointly, people will form a single integrated spatial representation of the town. However, if the neighborhoods are first learned separately and later observed to be connected, people will form separate spatial representations; this should incur an accuracy cost when inferring directions from one neighborhood to the other. Interestingly, our data instead suggest that people have a very strong tendency to form local representations, regardless of whether the neighborhoods were learned together or separately. Only when all visible distinctions between neighborhoods were removed did people behave as if they formed one integrated spatial representation. These data are broadly consistent with evidence from rodent hippocampal place cell recordings in connected boxes, and with hierarchical models of spatial coding.