edit: More directly answering the initial question, thank you, I misread what you wondered and was mislead a bit by a previous comment and your response to that comment.
In these experiments, they are using optical, not electrical techniques. Therefore it is a bit difficult to estimate exactly the number of cells activated directly, because light will scatter in the tissue but also fall off with distance. Based on the density of labeled and affected cells (to be stimulated, the cell has to be both a) labeled by a virus injection, and b) active in the engram) pictured in the second link you included, it seems like no more than a few to a couple dozen cells would have been activated directly.
In your comment, you said "IOW, how finely can implanted electrodes read and stimulate neurons?" - this is a different question. Implanted electrodes can read from and stimulate single neurons, or populations, depending on the electrode and experiment. It is difficult to target particular neurons with electrodes without an additional optical technique.
In the experiments you reference, optogenetic tools are used to selectively label cells that were active earlier in an experiment, and then reactivate those same cells with light. These sorts of experiments have only become possible recently with the development of those tools.
(original answer below)
Standard answer of biology applies here: "it varies", probably by species, type of memory, brain structure, etc.
But, to give you some ideas, one area that has been studied a lot in this field is the dentate gyrus of the hippocampus (the second and third links you sent, for example). One paper showed that a given memory activated something like 2-4% of the cells in that structure in a mouse http://www.sciencedirect.com/science/article/pii/S0166432811006462 with something like 1,000,000 dentate gyrus cells total in a mouse. So, that's a lot of cells, but those cells might be doing something special (pattern separation), so maybe it isn't right to think of them as the "engram" itself, or maybe we should be counting them PLUS all the cells activated in CA1, PLUS all the cells activated in cortex, etc. Whatever the case, one thing that is clear is that cells are involved in many many different engrams - if you are familiar with pop neuroscience you might recall the excitement about "grandmother cells" and "Jennifer Aniston" cells from a few years back. That's almost certainly not how engrams work, but that out of the subset of neurons and stimuli they examined, there was only limited overlap in those cases.
For fMRI studies, it's pretty hard to get a handle on the number of cells involved, but it isn't really possible to measure from "small clusters of cells" in fMRI on the same scale as in other measurement techniques. What is going to be more important for the fMRI decoding studies is "variation within a brain region" where the total number of neurons is many many millions, but the exact number involved in any one engram can't be measured with current technology.
