CAMELoT: Towards Large Language Models with Training-Free Consolidated Associative Memory