The headline is misleading because it sounds like "one neuron per memory". I will read the Nature article tomorrow, but in the reporting I see nothing at all about how many neurons are involved. Presumably the "memories" are distributed across many neurons.

It appears that what they actually accomplished was to find a way to evoke a conditioned association by shining light on the part of the brain that learns, rather than by reproducing the original stimulus. And this works only because they have genetically engineered the mouse so that the neurons which learn also become optically triggerable.

What MIT researchers MAY have demonstrated is that by "triggerering" a group of neurons (neural network) one causes the recall of a percept or series of percepts from an atemporal repository of memory, not from the group of neurons. Think of the complexity and depth of a memory you can recall from childhood and the speed with which you recall it. To say that the memory is stored in the neural network (or worse, a single neuron) is just silly. The recalled memory causes the individual to re-experience the image/event through the same neural network that forwarded the memory to the atemporal depository. Remove that neural network and you will not be able to recall and therefore re-experience the memory. Since it is totally impossible to store a memory or even a single percept in a neural network, what the MIT researchers MAY have demonstrated is the relationship and function of neural networks in the brain relative to the true repository of memory (capable of such storage and speed of recall) the atemporal component of all cells. See also, The Truly Astonishing Hypothesis at http://www.astonishinghypothesis.com/author/hypothesis.htm.