1. In schema mode, is the working memory schema included in the context window alongside each user message? Or id it more like a tool that can be called? 2. Is the full working memory JSON object also included in the content window with each user message? Or does the agent read from the working memory only when it needs to, just like any other tool use? 3. If the working memory is scoped to an agent (which is my understanding) and to a resource (my use case), do (sub)agents-as-steps also write to their own working memory or to the parent working memory? 4. Do you have any examples of clients using super-detailed working memory schemas successfully? I feel working memory is a bit unlock for me and I'm probably getting too exciting about it! Like, why would I not want to have a large, well-structured working memory for every user (that goes beyond the todo-list/user information examples you have in your docs)?
Side note, your docs are EXCELLENT. I also loved the MCP course.