Based on the sensory information received from the environment, each agent can build its own world state. We developed a predictive memory model that builds a probabilistic representation of the state based on past observations. By making the right assumptions, an effective model can be created that can store and update knowledge even when there are inaccessible parts of the environment. The agent relies on past observations to determine the positions of objects that are not currently visible. We conducted experiments to compare the effectiveness of this approach with a simpler approach, which ignored the inaccessible parts of the environment. The results obtained demonstrate that this predictive approach does generate an effective memory model, which outperforms a non-predictive model .