Well, your current architecture solves the problem of large context. But still it perform like openai's mention. Wher we have to manually set a flow of actions.