1 article on this topic.
RLMs treat long context as a REPL, not a stuffed window. Let the agent write Qdrant queries, then inject tenant filters before search.