The simulation only needs to produce observations that are consistent with the knowledge of the first observer. Sometimes bit even that, as I describe in the blog post, because eyewitness testimony is known to be quite unreliable.
I'm not sure what sort of history you're thinking of specifically.
The amount of information science could infer on many questions is strictly bounded and in those cases we could only reason stochastically. The data presented at time of first observation can then be generated randomly from the set of answers consistent with what's already known.
The place to implement an optimization like this would be at the items with the biggest O in your world, which is usually the smallest building block--what you'd have the "most" of, which would drive the largest memory and processing demands.