Meta Rolls Wonder-Trained Specialized Models into Muse and Meta AI: Active Memory/Focus; Vendor PSI ~38%→47%

Core Background & Industry Pain Points

Personal agents often fail less on single answers than on forgetting goal-critical context or freezing when plans change. Weeks ago Meta showed Active Memory and Active Focus on Muse Wonder. On 2026-11-06 it rolled specialized models trained with Wonder into Muse and Meta AI so more people get some of those frontier behaviors without waiting for Wonder as the default.

Architecture Highlights & Internals

The specialized models inherit two Wonder capabilities: Active Memory separates temporary from permanent facts, recalls earlier details when they matter again, and marks stale information; Active Focus recovers or tries alternatives when bottlenecks appear. Meta also highlights better multi-tasking and mid-task steers in Muse, and sharper step-breaking / follow-ups / direction changes in Meta AI. Wonder itself remains for harder work.

Authoritative Benchmarks & Measured Scores

Figures below are Meta research-blog / @AIatMeta vendor-reported, not reproduced here. Personal Superintelligence Index ~38%→47%. Active Memory Eval ~76%→90%. In Muse, forgetting goal-critical information drops about 2.4×; finishing everyday tasks amid unexpected constraints improves about 67% (exact baseline not detailed). No public head-to-head memory board vs Claude/GPT was published.

Developer Hands-on Guide

Muse / Meta AI users typically need no settings change. For memory agents, write acceptance tests around temporary-vs-permanent, staleness, and recover-or-reroute. Until Meta opens Memory/Focus eval sets, treat PSI / Active Memory numbers as vendor ceilings and instrument your own forget/unstick rates.