Meta launches Muse Wonder: what its memory and focus mean for AI agents

Watch: Meta’s Muse and AI memory

Watch Promptwire’s video on Muse, then explore its memory, focus and practical limitations below. Watch on YouTube.

What is available now—and what is coming later

Meta’s announcement lists Muse Wonder through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Integration into the Muse personal agent will follow, beginning with subscribers. A model release and an app rollout are separate events.

For anyone evaluating it from India, confirm eligibility and availability in the actual service before planning a team rollout. This article does not establish universal country access or quote unverified subscription and API prices.

Public record: Meta Newsroom — Muse Wonder announcement, October 12, 2026

Active Memory: carrying useful context forward

Meta describes Active Memory as retaining important context across conversations and retrieving it when relevant. Its research post frames this as a requirement for useful personal and professional agents.

For an agency, our illustrative test would use a fictional client brief with a target audience, tone and approved offer. Change the offer in a later message, then ask for a campaign draft. Check whether the new offer replaces the old one and whether the output preserves the remaining constraints.

Remembering more information is only useful when the agent selects the right version. A team should test updated facts and conflicting instructions, rather than checking recall only on an unchanged brief.

Public record: Meta Research — Muse Wonder capabilities and evaluation overview

Active Focus: continuing when work becomes difficult

Meta says Active Focus supports trying alternative approaches, recovering from blockers and adapting as a task develops. Its evaluations include FocusBench and SWE-Marathon, covering persistence and sustained software work.

Our proposed test: give the agent a small internal prototype brief, then introduce a missing requirement midway. Inspect the final deliverable against a checklist written before the run. Count how often someone must restate the goal, correct an assumption or repair the output.

Persistence should improve completion quality. An agent that continues confidently after misunderstanding the task can increase review time. Include a point where asking a precise question is the correct next step.

Public record: Meta Research — Muse Wonder capabilities and evaluation overview

Capabilities and benchmark limits

The model card lists text, image, video and audio inputs, text output and a one-million-token context window. It also warns about inaccurate responses, multi-step errors, spatial reasoning limitations and uneven language performance.

Meta’s evaluation methodology mixes internally generated and externally reported results. Some comparisons use different agent setups; its internal competitor runs are described as best-effort. These details matter when interpreting a claim of superior performance. Promptwire has not independently tested these models.

A large context window does not by itself prove dependable memory, and a benchmark ranking does not establish fit for a client workflow. Evaluate the combination of model, tools and instructions you would actually use.

Public record: Meta Research — Muse Wonder model card

Three business pilots worth considering

The following are editorial suggestions, not measured product results. Use fictional or approved sample material and inspect each deliverable before expanding a pilot.

Compare against your current workflow using the same inputs and acceptance criteria.

Keep memory and permissions in the evaluation

Meta’s model card recommends application-specific evaluation and added safeguards for agents that take real-world actions. It identifies human confirmation for irreversible actions as one example.

For a business pilot, decide what the agent may read, what it may remember and what it may change. Separate preparing a draft from sending it. Record actions so a reviewer can see how the agent reached the final state.

Also test removal and correction of remembered information in the product you choose. The launch description alone does not demonstrate how every memory control works in each deployment.

Public record: Meta Research — Muse Wonder model card

The Promptwire take

The useful question is whether an agent can carry a changing brief through to acceptable finished work. Muse Wonder gives teams another candidate for that evaluation.

Start with one recurring task. Measure completion quality, human intervention, elapsed time and actual charges. If the output still needs substantial repair, a more persistent agent has not yet improved your workflow. Adopt on the strength of those results.