I see some great ideas here already. One other thing to consider is the Roux method, which I think requires a smaller amount of memorised algorithms. I learned the usual layer-by-layer method years ago and tried switching to Roux a year ago, and it’s fun, although I don’t think I’m faster since I don’t really practice (maybe 45-60 seconds with layer-by-layer, and around a minute for Roux).
When doing layer-by-layer, I did come up with a mnemonic system for learning algorithms: each face move translates to a digit, then putting 7 before a digit means to turn anti-clockwise, putting 8 before means turning twice, then 9 and 0 signify slice and wide moves. I had some other method to indicate cube rotations. This would turn each algorithm into a sequence of digits which I could memorise with the Major system. But it became a real chore for me, and didn’t help with speed since you want to build cube algorithms into muscle memory without having to think about them.
Another possibility is to use the blindfolded methods. I use the most beginnerish Old Pochmann method which is really only 3 algorithms: one for edge swaps, one for corner swaps, one to fix parity errors. During the solve, you just have to memorise one sequence of letters for the order of edge swaps and another for corner swaps, then intuitively figure out how to setup (move the target piece to a “buffer slot” position) and restore (undo the setup moves). I use a memory palace for the piece letters: edge pieces go in the sitting room, corner pieces in the bedroom.