Best practice would be to either add more physical RAM or decrease the allocation to each VM. 64 GB of RAM in am ESXi host is not much to begin with.
Most operating systems will touch all memory pages on boot of the OS, which means that if you have 1 host with 96GB of memory allocated across 3 VMs but only 64GB of physical memory backing it that you are indeed oversubscribed. At this point you may see memory reclamation techniques kicking in even.
You should monitor the VMs and first and foremost right-size them! If they use 8 or 12GB then you assign them what they use plus some headspace.
Next you can consider enabling TPS across VMs and disabling Large Pages. This should allow TPS to "dedupe" similar memory pages and should lower the memory consumption.
If all of that fails, then you simply should buy more memory