You would have to one, make the procedural math solid. Put in the error-checking routines so that nothing comes undone, numbers-wise. And B, have a way to save those numbers in a system that would know they are no longer relevant but still important to know if they should become relevant again.
In other words, you have your randomly generated star systems all around you, you visit them all, they are all generated and running. You travel billions of lightyears away and visit another randomly generated star system. You now have two randomly generated star systems in memory, the one you are closest to is relevant and should be up and running. The other one should be in storage, not really running. The system should load the further system when the further system is no longer further away, calculate the differences in time for proper orbital alignments and such, making sure all of the space debris and soil sample caches are back in their proper place considering the amount of time past.
Sounds like you’re Describing Noctis. Noctis was a dos based star system exploration game with poor blurry VGA like graphics but it had a vast star mapping system. Meanwhile, No Man’s Sky did have pretty graphics, but that game just 1) kept on crashing out to desktop when you try to explore around the planets whenever the game hit an error in its calculations, and the game was also stripped down of over half its features, so half the features were already missing out of the public release.
Because on each planet you visit, you kept on coming up wtih the same alien model characters no matter how many zillion of worlds that you visit, all that changed was the shop names of the alien traders, but not the models or the buildings, because the same alien model keeps on popping up all the time in each trading store you visit of each planet so there was very little substance even though you had millions of worlds with all kinds of different wildlife.
However, my game has only got about 120
systems in it. Not millions of systems.