We need to start making benchmarks in memory & continued processing over a task over multiple days, handoffs, etc (ie. 'agentic' behavior). Not sure how possible this is.