A proper system needs to store the original locations and rotations, read them back maybe 3 ahead, and do the math over what it needs to do to achieve said goal.
You could store input pressed + the transform. That way you have some idea of what action was being performed (or rather starting) at the time.
But still, if you want exact behaviour, you may want to essentially record the animation that took place.
Sequencer does that at the editor level.
You may be able to use that as part of gameplay.
If not, then you can draw inspiration and extract code from it.
Each frame should record the transform of each bone making it possible to play back the exact same motion.
Actually, im sure this is part of the gameplay. Look up how to do rewind time systems. I believe its the same thing in practice.