Root cause: recently-shipped memories are backward-looking. Without a forward-looking ordered list, each new session started from zero and picked whatever felt salient.
Fix: mem://preferences/roadmap holds an ordered next-move list ranked by impact ÷ effort, plus a do-not-do list. Core memory instructs agents to read it before planning.
Prevention: any recurring question the agent answers from scratch is a signal that the answer belongs in memory, not in chat.