The canonical reference for Streams is in SICP, where Abelson and Sussman introduce the concept in Section 3.5. The motivation behind Streams is to model a list of potentially infinite size. Clearly, even if you take infinite to mean "a very big list", memory would be a problem. In the example from SICP, they consider the case of searching for large prime integers. It would be more schemely to use something like filter on a list of integers to return the list of prime integers for a given range, but for large integers, you would need to retain a very long list, because primes get farther apart the bigger the numbers in question get. The idea extends to any data structure that's bigger than memory, or to any operation that's expensive (either computationally or due to I/O costs) - delaying the evaluation of any expensive operation will improve performance, particularly if the results of evaluating that expression may not be used.
SRFI-41(replaces SRFI-40) defines a portable stream library. The primary difference between the SRFI and the SICP approaches is that SICP introduced an "odd" stream approach, while the SRFI follows an "even" approach. The SRFI document has an excellent comparison of the two approaches, but the simple explanation of the difference is that even streams delay evaluation of the stream until absolutely necessary, whereas odd streams are evaluated at the time of reference, and are therefore subject to being evaluated once more than necessary.