Cool! Finally managed to explain it in terms that make sense to someone else…
The time-quantisation would be optional, but I think it has to loop, otherwise there’s no real advantage over interpolated seeded random sequence.
I don’t think you’d ever have to save anything in RAM, except the lookup table used to store the grid values used to lookup the perlin noise. This would be a constant though, and would never have to be written to at runtime.
Most implementations I’ve seen use a largish array (typically 256 or 512 random values).
My implementation wouldn’t ever need to pre-generate a sequence of values. It simply looks them up at each point, so the computational load is the same for every lookup. The downside of course, is that that load might be quite high, depending on the interpolation method used, so it might not be possible to run at audio rates. For me, that’s not a problem.