Since this post was made I discovered double-batched Merkle accumulators which are better than MMRs, and are much easier to understand. The scheme above works the same if you replace the accumulator.
If you’re still interested in MMRs (a.k.a. asynchronous accumulators) you can find a visual depiction of the algorithm in pages 18 and 19 in this paper. An MMR is basically like a binary counter where each bit is replaced with a Merkle tree root, and the carry operation merges Merkle tree roots.