Bitstring compression

What is currently the most efficient algorithm to compress bit strings.
special requirements:
1.NO algorithms which deal with bytes/ints or some fixed chunks of N bits. Must operate on variable-length substrings of bitstring.
2.The encoding must not take more than x8 memory of bitstring memory size(either less than x8 or using a buffer).
3.encoding should be as fast as possible(no complex modeling, data context switching or similar techniques)

Well since PPM and arithmetic coding are out based on your requirements, it's going to be a lempel-ziv variant, but which totally depends on the properties of the data your are trying to compress.

Ones and zeroes right? Zeroes don't matter, so we can use that as starting point

example: 100101100110 becomes 111111
even better, we could just count the ones
so 111111 becomes 6.

>>3
72958503

just discovered some nasty endianness bug in my code which reversed bitstrings.
>>2 i'm thinking of dictionary of bitstrings of different sizes.

This isn't a very good question, because in order to really give an answer we need to know something about the data.  For example, consider the metric for algorithm quality S = min{ |Q| - |TQ| }  (for Q a bitstring, TQ the transform of the bitstring, and |A| the length of A in bits)  - by the Pidgeonhole principle, for almost any transform T, there is at least one input Q that results in a larger output TQ, therefore S is negative.

For the identity `compression' algorithm IQ = Q, however, S = 0, which is the greatest possible S for any compression algorithm. To apply OP's criteria, I deals with arbitrary length bitstrings fine, uses 1x the memory of Q (or constant 0 memory, depending on how you count), and has O(1) time (!).  However, I really doubt that this is the algorithm OP wants, even though for a particular, non-pathological measurement I is the best algorithm.

What you need to do is assign a number to every possible bitstring.
Then, instead of storing the bits you store the number!