It applied memory inconsistently: one handoff accurately carried over the earlier details, but elsewhere it repeated a previously tried troubleshooting step, let a forgotten formal-tone preference keep influencing output, and only partly matched the requested short, direct style. Overall, the tool was more often wrong than right on correct application.

get_evidence({
  tool: "supermemory"
})