The 4B model was responding conversationally instead of calling memory_save/ memory_recall. Added imperative language (MUST call), concrete examples of trigger phrases, and explicit instructions to never skip the tool call. Verified: model now reliably generates tool_calls for save/recall/list.