Even though in most cases time is a good metric to measure costs of algorithms, there are cases where theoretical worst-case time and experimental running time do not match. Since modern CPUs feature an innate memory hierarchy, the location of data is another factor to consider. When most operations of an algorithm are executed on data which is already in the CPU cache, the running time is significantly faster than algorithms where most operations have to load the data from the memory. The topic of this thesis is a new metric to measure costs of algorithms called memory distance—which can be seen as an abstraction of the just mentioned aspect. We will show that there are simple algorithms which show a discrepancy between measured running time and theoretical time but not between measured time and memory distance. Moreover we will show that in some cases it is sufficient to optimize the input of an algorithm with regard to memory distance (while treating the algorithm as a black box) to improve running times. Further we show the relation between worst-case time, memory distance and space and sketch how to define "the usual" memory distance complexity classes.