Interesting approach. Curious how you think about the tradeoff between external memory systems vs. models that manage their own context natively. It feels like the industry is split on which layer should own this.