Each feature dataset has a XY tolerance that is defined during its creation, and cannot be changed. The tolerance of a topology can be changed IF the geodatabase is not registered as versioned.
When you create a topology, you can work with more than one layer. So you have a set of feature classes with (potentially) different XY tolerances. In order to determine the common rule for coincident points, you therefore need to set one XY (=cluster) tolerance value for your topology. The default value for the tolerance in a topology is 10*tolerance of dataset. And the default value of XY tolerance in a dataset is 1 millimeter.
e.g. A has a tolerance of 1 and B has a tolerance of 10, so the coordinate point (6,6) can be stored in A, but if you try to store it in B it will be rounded to (10,10). If you select 1 as a cluster tolerance for topology, those points will not overlap; If you select 10, they will.
EDIT: more details can be found here, but, in short, you should relate the value of your tolerance to the value of the precision of measure.