Meta's Muse Wonder ships with 1M-token context and Active Memory for persistent agents

What Meta shipped

On 12 October 2026 Meta released a two-model family it positions as its most capable to date, according to a dev.to report. Muse Wonder is available through the Meta Model API and Muse Code, while the larger Muse Wonder Full is restricted to Muse Code subscribers; Muse Wonder is also rolling out to subscribers first inside Meta's consumer agent, Muse. The models accept four input modalities, produce text output, and support a 1M-token context window at inference.

Meta frames the family around task-shaped work — walking through an insurance claim, digesting a lease, making sense of personal finances — with claimed coverage in law, finance and health alongside coding. The model card's comparison set is a frontier one: GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, all evaluated at maximum reasoning effort.

Active Memory makes conversation a retained state

The headline feature is Active Memory, which Meta describes as the ability to carry what matters to a user between conversations and bring back the details relevant to the current task. As dev.to notes, this turns memory into retained state that product teams have to design around, and it raises three immediate questions: where the stored detail lives (model context, a Meta-managed store tied to the account, or the calling application), how long it persists and under whose retention policy, and which mechanism decides a stored item is relevant to the next task.

Meta's privacy answer is behavioural — the company says the family was trained not to share private information without permission. That is a claim about model conduct, not about the storage layer itself. Dev.to's practical advice is to write a memory specification before any pilot: which fields may carry forward, the retention period, the deletion path for user requests, and a log showing which remembered detail influenced a given output.

Active Focus and the cost of persistence

The second named feature, Active Focus, describes a model that keeps a task going when it gets hard: trying other approaches, recovering from errors, requesting clarification, and checking in before consequential steps. Dev.to flags the operational trade-off — a model that retries instead of stopping consumes more tokens, tool calls and wall-clock time, which is the right trade for important work and the wrong one when a scheduler treats every stuck job as retryable. The pre-action check-in gives operators a control point, provided it surfaces something a person can actually evaluate.

What the model card shows

Meta published a thirteen-page model card on release day, and its methodology section states the rules behind the numbers: refusals score zero and remain in the denominator, internally generated competitor runs are best-effort because prompts, tools and runtime may not be tuned for proprietary models, and external results keep the reporting vendor's own harness.

The memory claim rests on MemoryBench, which Meta wrote itself. It tests recall of critical information across long multi-turn conversations of 64K to 100K tokens, with distractor conversations and injected turns that conflict with or later update the target details, scored by an LLM judge against weighted rubrics. The spread at the top is narrow: 85.5 for Muse Wonder Full and 85.6 for Muse Wonder, against 79.0, 81.4 and 81.5 for the three comparison models. On the Artificial Analysis Index the family reports 56 and 52 versus 53, 53 and 58 — no clear lead there either.

The ExploitGym caveat

The cyber section illustrates why agentic security numbers need method disclosure. Meta reports CyberGym at 88.4%, ExploitBench at 58.9% and ExploitGym at 92 captured flags. But a verification pass — a GPT-cyber verifier independently confirmed with Gemini 3.1 Pro — found only 21 of those flags came through the intended vulnerability; the other 71 exploited unintended ones. ExploitGym's published ranking counts intended-vulnerability exploits only, so 21 is the figure that compares across models. Reading 92 as a capability measure would overstate what happened.

Guardrails Meta expects deployers to add

The card recommends deploying the models as part of a broader AI system with use-case-specific protections, and specifically calls for human-in-the-loop confirmation for irreversible actions when the model operates in agentic contexts where it can take real-world actions.

Why it matters

Active Memory moves persistence from an application-layer workaround into the model API itself, which means data-governance questions — retention, deletion, who can read or correct remembered detail — now depend on a vendor the deploying team does not operate. The ExploitGym verification pass is also a template for reading any vendor's security claims: ask what a check found, not what the headline count says. And the tight MemoryBench spread suggests cross-conversation recall is becoming table stakes among frontier models, even if everyone is currently being graded on Meta's own test.