You doing pure BP AI, no behavior trees or EQS?
If so you should dig int those 2, behavior trees are not that great. Same functionality may be done with about same time investment compared to BPS, that if you count time needed for learning trees and coding it.
But huge benefit of behavioral trees is EQS. For your code here, you could run EQS query, that finds best spot to run to that was near last seen location, store this in btree key as vector.
That stuff is instanced per aicontroller, so no problems about which bot should remember which spot.
Downside of it is that atm EQS is a bit wonky, but i think they will patch it soon enough.
For memory (in pure blueprint solution), you should store vector of last location.
Or even better create struct, that stores vector, forward vector of enemy when last seen (to remember where it ran), and maybe some other stuff that you possibly want to add later (that s major reason for struct here, you can expand it later without big changes to code).
Store that struct in your player controller of that bot. Better even make array out of it, tore all those points, then make some filtering rule and pick best (again EQS is very powerfull for this).
For bots to behave like bots you need to thing like that animal or bot would think, and code in similar way to their decision process, this way bots will behave in more “natural” way. So storing last seen location in local bot memory is more natural than some global invisible actor.
Ps. you could also add expiration data for last seen location. So bots lose interest after some time, but that was very weird in SKYRIM for eg. After you stealthy kill somebody his buddy who was chatting with dead now since forever, say “Oh probably that was nothing”, 10 seconds later “oh no where is that body from!” about guy they were chatting all the time since beginning of game. I find it creepy.