So if memories are being replayed, where are the memories coming from and to where are they being persisted?
Is it like a backup is being formed from more short-term memory to more long-term memory? Are long-term memories being replayed to ganglia responsible for "muscle memory" to improve response times?
Recent memories (so-called 'hippocampus-dependent memories') are stored in the hippocampus for up to about two weeks, during which time they are gradually transferred into permanent storage in the neocortex, cerebellum, and other parts of the brain. So, to answer your question, that is where they are replayed from.
The hypothesis is that the hippocampus is unique in being able to store memories quickly and after a single exposure. Most of the rest of the brain needs time and repeated exposures to store a memory via long-term potentiation. The role of the hippocampus is to act as a buffer where memories are kept temporarily.
While you're sleeping, your hippocampus repeatedly activates the memory, stimulating the brain in a similar way that the original event did. Slowly, the neocortex begins to rewire itself through synaptic plasticity to store the memory. After many repetitions, the memory is no longer dependent on the hippocampus and it fades away.
I think this is the paper I'm recalling this from, but I'm not entirely sure (this memory is several years old):
Bontempi, Bruno, et al. "Time-dependent reorganization of brain circuitry underlying long-term memory storage." Nature 400.6745 (1999): 671-675.
How could you test for something like that? Maybe, teach an array of skills, and for each skill with subjects of varying degrees of sleep deprivation and for varying time intervals between education and performance, measure response times and skill level in the tasks?
One way this 'replay' effect was tested was in rats where they implanted many electrodes into the animal's hippocampi to monitor activity. There are cells there called the hippocampal place cells which encode locations in space.
Then they had the rats run a simple circular maze with some landmarks (food, etc.) at various points. As they run the maze, different place cells start to encode different locations in the maze, especially near the landmarks, and they activate when the rat is near there.
They kept the electrodes in while the rat was sleeping. They found that the place cells continued to activate in the same order as the locations in the maze, as if the rat were actually running the maze in, e.g., a clockwise direction.
This is taken as evidence that, during sleep, the rat's brain was replaying the memory of running the maze.
Is it like a backup is being formed from more short-term memory to more long-term memory? Are long-term memories being replayed to ganglia responsible for "muscle memory" to improve response times?