Despite its relevance for navigation surprisingly little is known about how goal direction bearings to distant locations are computed. Behavioural and neuroscientific models proposing the path integration of previously navigated routes are supported indirectly by neural data, but behavioral evidence is lacking. We show that humans integrate navigated routes post-hoc and incrementally while conducting goal direction estimates. Participants learned a multi-corridor layout by walking through a virtual environment. Throughout learning, participants repeatedly performed pairwise pointing from the start location, end location, and each turn location between segments. Pointing latency increased with the number of corridors to the target and decreased with pointing experience rather than environmental familiarity. Bimodal pointing distributions indicate that participants made systematic errors, for example, mixing up turns or forgetting segments. Modeling these error sources suggests that pointing did not rely on one unified, but rather multiple representations of the experimental environment. We conclude that participants performed incremental on-the-fly calculations of goal direction estimates within compartmentalised representations, which was quicker for nearby goals and became faster with repeated pointing. Within navigated environments humans do not compute difference vectors from coordinates of a globally consistent integrated “map in the head”.