The simulator sends sensory information to the client about the portion of the state that is visible or audible. This information comes in two forms: visual and auditory. Auditory information is just a message that was spoken and from what direction the message came from. This provides little information about the current state since there's no simple way to determine who spoke the message.
Visual information is provided as a list of objects with their relative distance and angle from the client. Information is only provided for objects within a 60 degree arc from the direction the client is facing. Objects within this arc will be referred to as visible objects, because their exact position is given by the simulator. Objects that the simulator provides information about are the ball, other players, the two goals, the four flags at the corners of the field, and the flags on the sides at mid-field. For players, the amount of information describing which player varies on how far the player is away. If the player is close then both the team and number is provided. If the player is further away then only the team is provided, and if very far away then no information is provided.
This is a very limited amount of information, making this domain highly inaccessible. There are many situations when most of the current state is not available to the client. If there were no memory model at all, then just turning away from the an object would cause the client to completely forget where the object was. For example, if the client receives a pass and then turns to look for the goal, it no longer knows where the ball is and won't be able to shoot. A memory model is essential for overcoming the inaccessibility of the environment.
Version 2.x of the simulator provides additional information to the client. At each step, the distance to the boundary directly ahead, along with the relative angle of the client's facing to that boundary line, is provided to the client. This information is in addition to the information described above. The memory model outlined in this paper equally applies to the newer simulator. The appendix provides calculations which use this added information to improve the memory model.