A whitepaper and implementation is available at https://github.com/tromp/cuckoo
From the README file:
Mining is generally considered to be inherently power hungry but it need not be. It’s a consequence of making the proof of work computationally intensive. If computation is minimized in favor of random access to gigabytes of memory (incurring long latencies), then mining will require large investments in RAM but relatively little power.
Cuckoo Cycle represents a breakthrough in three important ways:

it performs only one very cheap siphash computation for about 3.3 random accesses to memory,

its memory requirement can be set arbitrarily and doesn’t allow for any timememory tradeoff.

verification of the proof of work is instant, requiring 2 sha256 and 42 siphash computations.
Runtime in Cuckoo Cycle is completely dominated by memory latency. It promotes the use of commodity generalpurpose hardware over custom designed singlepurpose hardware.
Other features:

proofs take the form of a length 42 cycle in the Cuckoo graph.

it has a natural notion of (base) difficulty, namely the number of edges in the graph; above about 60% of size, a 42cycle is almost guaranteed, but below 50% the probability starts to fall sharply.

running time for the current implementation on high end x86 is under 24s/GB singlethreaded, and under 3s/GB for 12 threads.

making cuckoo use a significant fraction of the typical memory of a botnet computer will send it into swaphell, and likely alert its owner.