{"schema":"sanctity-signal/0.2","record_type":"signal","id":"meta-active-memory-everyone-e078","domain":"labs","drop_id":"signal-e078","occurred_at":"2026-11-06","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Meta","action_type":"rollout of Muse Wonder-trained specialized models carrying Active Memory and Active Focus to all users of Muse and Meta AI, with in-house evaluation scores published","observed":"Meta published Bringing Specialized Models to Muse and Meta AI at research.meta.ai on November 6, 2026, and we read it there. Mark Zuckerberg posted about it on X at 18:08 Universal Time and the AI at Meta account at 18:14 Universal Time the same day. What ships: specialized versions of Meta's existing models, trained using Muse Wonder, rolling out to everyone who uses Muse and Meta AI. The page does not name the specialized models; its charts refer only to Current Muse Model and Previous Muse Model. Muse Wonder itself stays reserved for harder subscriber workloads. Active Memory is described as helping the assistant keep track of what matters and recognise when earlier information becomes relevant again, with the models trained to tell temporary information from permanent, reason across time, and separate what is still relevant from what is stale. Active Focus is described as letting the assistant keep pursuing a goal when unexpected obstacles appear, trained on real-world-style bottlenecks so the models try harder, recover, or find alternatives. The page says the models also handle multiple tasks, decide when to push on goals independently or collaborate live, and adapt to mid-task steers, and describes the same improvements in Meta AI as breaking problems into steps. The 4 numbers, all Meta's own: on the Personal Superintelligence Index, the current model scores 47% against the previous model's 38%; on Meta's Active Memory evaluation, measuring tasks correctly using critical context, 90% against 76%; a 2.4 times reduction in forgetting critical information necessary for helping you achieve your goals; and a 67% improvement in completing everyday tasks where unexpected issues arise. The page names a baseline for the first 2 figures, the previous Muse model, and gives none for the other 2, nor any description of how they were measured. Zuckerberg's post, as quoted by Digg, says Meta is bringing Active Memory and Active Focus from Muse Wonder to everyone. Not established: the identity or size of the specialized models, any outside evaluation, separate figures for Meta AI, the regions or account tiers reached first, and what the Personal Superintelligence Index measures.","corrections":[],"evidence":{"status":"verified","note":"Verified. We read the announcement at research.meta.ai, which carries the 2 scored comparisons, the feature descriptions and charts labelled Current and Previous Muse Model. Its metadata gives November 6, 2026 at midnight, so the blog gives no usable time; Mark Zuckerberg posted at 18:08 Universal Time and AI at Meta at 18:14, which places the announcement in this window.","sources":[{"role":"primary","label":"Meta AI Research: Bringing Specialized Models to Muse and Meta AI, Nov 6, 2026, read at research.meta.ai","url":"https://research.meta.ai/blog/specialized-models-for-muse-and-metaai"},{"role":"corrob","label":"AI Weekly: Meta ships Muse Wonder-trained models to all Muse, Meta AI users, Nov 7, 2026, 02:02 UTC","url":"https://aiweekly.co/alerts/meta-ships-muse-wonder-trained-models-to-all-muse-meta-ai-users"},{"role":"corrob","label":"Digg: Meta announces Active Memory and Active Focus upgrades for Muse and Meta AI, Nov 6, 2026","url":"https://digg.com/ai/vc7xw9zf"},{"role":"corrob","label":"Technobezz: Meta Brings Active Memory and Active Focus to Muse and Meta AI, Nov 2026","url":"https://www.technobezz.com/news/meta-muse-meta-ai-memory-focus"}]},"strength":{"cost":"Medium","commitment":"Medium","reversibility":"Moderate","persistence":"Year","rationale":"Cost is medium: training specialized derivatives and serving them to a consumer base of this size is real money, and giving away a subscriber feature forgoes revenue, but Meta keeps its frontier model behind the paywall, so the giveaway is bounded. Commitment is medium rather than high, because a rollout can be slowed, staged or reversed without filing anything, and the page names no schedule. Reversibility is moderate for the same reason, although memory is the hardest feature to withdraw once users rely on it. Persistence is a year: this is a product generation, not a structure."},"shift":{"prior":"Persistent memory and goal-chasing were a paying subscriber's feature.","action":"Meta published Bringing Specialized Models to Muse and Meta AI at research.meta.ai on November 6, 2026, and we read it there. Mark Zuckerberg posted about it on X at 18:08 Universal Time and the AI at Meta account at 18:14 Universal Time the same day. What ships: specialized versions of Meta's existing models, trained using Muse Wonder, rolling out to everyone who uses Muse and Meta AI. The page does not name the specialized models; its charts refer only to Current Muse Model and Previous Muse Model. Muse Wonder itself stays reserved for harder subscriber workloads. Active Memory is described as helping the assistant keep track of what matters and recognise when earlier information becomes relevant again, with the models trained to tell temporary information from permanent, reason across time, and separate what is still relevant from what is stale. Active Focus is described as letting the assistant keep pursuing a goal when unexpected obstacles appear, trained on real-world-style bottlenecks so the models try harder, recover, or find alternatives. The page says the models also handle multiple tasks, decide when to push on goals independently or collaborate live, and adapt to mid-task steers, and describes the same improvements in Meta AI as breaking problems into steps. The 4 numbers, all Meta's own: on the Personal Superintelligence Index, the current model scores 47% against the previous model's 38%; on Meta's Active Memory evaluation, measuring tasks correctly using critical context, 90% against 76%; a 2.4 times reduction in forgetting critical information necessary for helping you achieve your goals; and a 67% improvement in completing everyday tasks where unexpected issues arise. The page names a baseline for the first 2 figures, the previous Muse model, and gives none for the other 2, nor any description of how they were measured. Zuckerberg's post, as quoted by Digg, says Meta is bringing Active Memory and Active Focus from Muse Wonder to everyone. Not established: the identity or size of the specialized models, any outside evaluation, separate figures for Meta AI, the regions or account tiers reached first, and what the Personal Superintelligence Index measures.","update":"They are the default for a consumer assistant, graded only in house."},"inference":{"established":["Meta shipped Muse Wonder-trained specialized models to all Muse and Meta AI users.","Meta reports 90% against 76% on its own Active Memory evaluation.","Meta reports 47% against 38% on its Personal Superintelligence Index."],"unknown":["That the gains hold outside Meta. No third party has evaluated these models.","What a 2.4 times reduction in forgetting means. Meta gives no baseline.","Any figure for Meta AI alone, or the rollout schedule by region."]},"narrative":{"hook":"Meta gave every Muse and Meta AI user models that remember and keep pushing.","plain":"Meta published scores for the new models against its previous ones. Caveat: every number is Meta's own, and 2 of the 4 come with no baseline at all.","story":["Memory is the feature that makes an assistant hard to leave: it is the switching cost. Meta is giving it to its whole user base rather than selling it, while keeping Muse Wonder for subscribers. The grading is the catch. 47% against 38% is a real gap on an index Meta wrote itself."],"trail":[{"d":"OCT 12","t":"Meta ships Muse Wonder, the model behind its agent, built to remember you. Edition 049's labs card, and the parent of this one."},{"d":"OCT 15","t":"Meta publishes 816 billion parameters of model weights, free to download and build on. Edition 052."},{"d":"OCT 21","t":"A New Mexico court orders Meta to pay $25 billion for lying about data control, a judgment that bears directly on an assistant built to retain what users tell it. Edition 058."},{"d":"OCT 30","t":"Britain's competition regulator opens an abuse case over Meta's terms for using publishers' content. Edition 067. Meta later pauses its November 20 commercial scraping terms indefinitely."},{"d":"NOV 6","t":"Meta ships Muse Wonder-trained specialized models to all Muse and Meta AI users, with Active Memory and Active Focus, and publishes 4 in-house figures, 2 of them without baselines."},{"d":"Reacts","r":true,"t":"Every Muse and Meta AI user whose assistant now retains details between conversations; Meta's paying subscribers, whose distinguishing feature just became free; OpenAI, Google and Anthropic, which charge for memory; the New Mexico court and Britain's competition regulator, both examining what Meta does with what it holds; and the evaluation labs with no access to the models being scored."}],"glossary":[{"term":"Specialized model","def":"A smaller model trained using a larger one's output or methods, so it can be served cheaply to everybody while the big model stays reserved."},{"term":"Baseline","def":"The thing a claimed improvement is measured against. A figure with no baseline cannot be checked, which is why 2 of these 4 cannot."},{"term":"Personal Superintelligence Index","def":"Meta's own evaluation of its assistants. Meta publishes the scores, writes the tasks and sets the marking, with no outside referee."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Memory is the retention weapon","text":"An assistant that remembers you is one you do not get round to replacing.","abstain":false},{"id":"b","label":"A routine model refresh","text":"Specialized derivatives of a frontier model ship on an ordinary cadence.","abstain":false},{"id":"c","label":"Meta is buying usage with features","text":"Giving the paid capability away is how you move a consumer base at scale.","abstain":false},{"id":"d","label":"The scoreboard is the product","text":"In-house indexes let a lab declare progress with no outside referee involved.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":"meta-muse-wonder-e049","what_to_watch":[{"linked_interpretation":"c","direction":"strengthens","observable_event":"OpenAI, Google or Anthropic moves persistent memory or an equivalent persistence feature from a paid tier to its free tier, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"An outside evaluator publishes results on the shipped Muse models showing no improvement over the previous generation on memory or multi-step task completion, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"d","direction":"strengthens","observable_event":"Meta publishes the Personal Superintelligence Index tasks and marking, or gives a baseline for the 2.4 times and 67% figures, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"}]}