I habe come across turning on caching means the llm has a faint memory of what was in the cache, even to unrelated queries. If this is the case its fully unreasonable to share the cache, because of possibility of information leakage.