"Thoughts" isn't quite what is being lost. Instead, you've still got the neurons that understand how goals relate to motor operations, and you've still got the motor neurons, but they've lost the neurotransmitter "protocol" (the calibrated receptivity-weights on the synapses connecting them) that allows them to communicate.
For me, an analogy to speech helps: a baby learning to speak must learn not just language in a high-level sense, but at a lower level, they must "learn" (i.e. calibrate neurotransmitter receptors attached to motor pathways) for how strongly each muscle in their vocal cords should respond to a given low-level phonemic motor-intent-signal from the brain.
If you didn't talk for years, those calibrations would drift, and although you'd remember what it was like to talk, you'd have to "learn to use" (i.e. calibrate) your vocal cords all over again.
The interesting thing here, is that this "calibration" (or lack thereof) continues to cascade through the entire graph of neurons: think backpropagation in machine learning. Motor-neurons that have nothing to talk to lose their ability to talk; and then the motor-intent-recognizing neurons find the only thing they had to talk to (the motor neurons) are now silent—as good as dead—and so they no longer get reward-trained and lose their receptivity-calibration—and on and on, back until you reach whatever subset of neurons is still doing things that achieve "rewards."
Muscle memory is generally for referring to higher order things like blocking in a fight, swinging a bat, and other fast reactions that get better through practice instead of thought. Eg actions that your brain higher order plans but isn't thought by many to be consciously controlled. In fact when you think about a lot of these things consciously you mess up. Think throwing darts, shooting pool, etc.
For me, an analogy to speech helps: a baby learning to speak must learn not just language in a high-level sense, but at a lower level, they must "learn" (i.e. calibrate neurotransmitter receptors attached to motor pathways) for how strongly each muscle in their vocal cords should respond to a given low-level phonemic motor-intent-signal from the brain.
If you didn't talk for years, those calibrations would drift, and although you'd remember what it was like to talk, you'd have to "learn to use" (i.e. calibrate) your vocal cords all over again.
The interesting thing here, is that this "calibration" (or lack thereof) continues to cascade through the entire graph of neurons: think backpropagation in machine learning. Motor-neurons that have nothing to talk to lose their ability to talk; and then the motor-intent-recognizing neurons find the only thing they had to talk to (the motor neurons) are now silent—as good as dead—and so they no longer get reward-trained and lose their receptivity-calibration—and on and on, back until you reach whatever subset of neurons is still doing things that achieve "rewards."