They do keep context to a point, but they can't hold everything in their memory, otherwise the longer a conversation went on the slower and more performance intensive doing that logic check would become. Server CPUs are not cheap, and ai models are already performance intensive.