Meta’s Muse Wonder model card gives a more specific meaning to its Active Memory claim. The company’s internal MemoryBench tests conversations containing 64,000 to 100,000 tokens, with distracting exchanges and details that conflict with or later update earlier facts. An AI judge scores answers against weighted criteria. That makes the evaluation relevant to a familiar personal-assistant problem: carrying forward a corrected plan rather than confidently repeating the original one.
The card separately lists a one-million-token inference context window. That capacity figure and the memory evaluation measure different things; neither establishes how reliably the finished Muse app will remember a person’s changing circumstances over months. Meta also says performance varies with prompts, tools and the surrounding agent system. Its October 12 announcement, updated October 13, still describes Wonder-powered Muse experiences as coming soon, starting with subscribers. This is analysis of published model evaluations, without hands-on testing or confirmation of an account-level upgrade.
Editorial take: A useful memory check should include a correction. Give an assistant a harmless planning task, change one important detail later, then ask it to prepare the result again and identify which version it used. Check whether the revised date or preference carries through every part of the output. Remembering more information is useful only when the assistant can select the information that still applies.