Visual landmarks provide crucial information for human navigation. But what defines a landmark? To be uniquely recognized, a landmark should be distinctive and salient, while providing precise and accurate positional information. It should also be permanent, e.g., to find back your car, a nearby church seems a better landmark compared with a truck, as you learned the truck might likely move. To this end, we here investigate the learning of landmark permanency while treating permanency as a probabilistic characteristic for human navigation. Particularly we study the learning behaviour when exposed to landmarks whose permanency feature is probabilistically defined. We hypothesize that humans will be able to learn this feature and assign higher weight to more permanent landmarks. To test the hypothesis, we used a homing task where participants had to return to a position that was surrounded by three landmarks. In the learning phase we manipulated the permanency of one landmark by secretly repositioning it prior to returning to the home position. The statistics of repositioning was drawn from a normal distribution. In the test phase we investigated the weight allocated to the non- permanent landmark by analysing its influence on the navigational performance. The first experiment revealed the probabilistic nature of learning the prior of landmark’s permanency, the second experiment confirmed the results in modifying the statistic of the prior. The third and the fourth experiments showed that priors of permanency can be updated by experiences highlighting the capacity of adaptation to the environment.