Meta Muse Wonder & Full: Cross-Session Memory, Long-Task Execution, and Access Guide

Meta Announces Muse Wonder Series

On October 13, 2024, Meta's research team officially announced the Muse Wonder series. The base model, Muse Wonder, and the more capable Muse Wonder Full are now accessible through specific developer channels, but the consumer-facing personal app is still listed as "coming soon" and will roll out to subscribers first.

Active Memory: Cross-Session Context Retention

Meta calls the cross-session memory capability Active Memory . It aims to carry important user background — preferences, constraints, goals — into later conversations and actively retrieve relevant details when needed. For example, if a user previously stated a travel budget, dietary restrictions, and a preference not to wake early, a later request for a day itinerary should automatically incorporate those constraints without repetition.

Meta evaluates this with an internal benchmark called MemoryBench , testing 64K–100K token conversations that include distractors, conflicting information, and subsequent updates. The published internal results show the model's ability to recover key facts under these conditions.

Caveat: The benchmark does not prove memory persistence over years. Actual product design — how memories are stored, updated, deleted, and which personal data are readable — depends on product decisions and user permissions.

Active Focus: Continuous Long-Task Execution

Active Focus refers to the model's ability to keep pushing a multi-step task forward: trying alternative approaches, handling obstacles, and adjusting actions based on new information. A concrete illustration is writing a small program: after generating the first code version, the model must run tests, diagnose failures, edit multiple files, and re-verify.

Both models have been trained on diverse long tasks to reduce mid-task abandonment. However, the official documentation explicitly states that the model will ask clarifying questions when needed and confirm before important actions . This constraint means the model cannot act with unlimited autonomy, and not all tasks will succeed.

Capability Comparison

Active Memory — Problem addressed: Bringing relevant background into subsequent tasks. What to verify: Whether old information is outdated or wrong context is used.

Active Focus — Problem addressed: Continuing multi-step tasks after obstacles. What to verify: Whether the model drifts from the goal or requires confirmation.

Another internal benchmark, FocusBench , covers tasks in housing, travel, IT, and finance, emphasizing whether the model continues working when the environment changes or obstacles appear.

Actual execution still requires tools and a runtime environment. The model decides the next step; the product framework supplies browser, file system, command execution, and task state. After API integration, how cross-session data is stored, how tools are connected, and how permissions are managed remain the application's responsibility.

Access Points: Models vs. Entryways

As of October 13, the official announcement and developer documentation define the following availability:

Muse Wonder (Model) — Current access: Meta Model API, Muse Code

Muse Wonder Full (Higher-capability model) — Current access: Only Muse Code Power Usage and Maximum Usage subscriptions

Meta Model API (Developer calling endpoint) — Current access: Pay-as-you-go, requires API key; cannot call Full currently

Muse Code (Terminal-based coding agent) — Current access: Supports macOS, Linux, Windows; pay-as-you-go and subscription plans

Muse Personal App (Consumer product for personal tasks) — Current access: Wonder still "coming soon", subscribers first

The diagram separates models from entry points to avoid conflating "model release" with "personal app fully upgraded."

Pricing and Quotas

Wonder's API model ID is muse-wonder-1.0. Per million tokens: input $4, cached input $0.20, output $20. Long tasks invoke the model repeatedly, so total cost depends on cumulative usage, not per-token price alone.

Full is only available through Muse Code subscriptions: Power Usage $50/month, Maximum Usage $200/month . No standalone API token pricing for Full has been published; setting Wonder's inference tier to max does not turn it into Full.

Developer portal: https://dev.meta.ai/

Data mode: Standard mode does not use sent content for training Meta models. Muse Code subscribers default to Contributor mode (allowing training in exchange for higher quota), which can be disabled in settings.

Full model, plan, and data mode details: https://dev.meta.ai/docs/models

Practical Guidance

To build your own agent, start with the Wonder API.

For a ready-to-use coding toolset inside a project, evaluate Muse Code.

The personal app's Wonder integration is not yet live; a Muse Code subscription does not grant personal app access.

Access requires a Meta Account; availability and payment may vary by region. Mainland China registration, payment, and connectivity were not tested in this report.

Benchmarks and Evaluation Methodology

Meta published a 13-page model card listing specifications, evaluation methods, and limitations. PSI is an internal composite metric; MemoryBench and FocusBench are internal benchmarks — not a unified third-party leaderboard.

SWE-Marathon cites results from an Abundant AI report: GPT-6 Astra running in Codex, Claude Fable 5.1 in Claude Code, and Full using the xhigh inference tier. The model card explicitly notes this row is not a controlled comparison under a single agent framework and cannot support a claim of "overall victory."

The model card states both models accept text, image, video, and audio inputs, but output is text only . The 1M token figure refers to inference-time context length, not infinite memory.

An official demo generated a playable 3D kitchen game from an image, showcasing input understanding and code generation. Wonder's native output does not include images or video; the model card lists visual and spatial reasoning, especially 3D tasks, as known limitations.

Early community feedback: developer Anshu reported preferring Astra's game feel under the same prompt, but used different inference tiers for each model — an anecdotal, uncontrolled observation. A quota consumption bug was reported by Alexandr Wang; Meta stated it has been fixed and Muse Code usage reset, though this fix has not been independently re-verified.

The model card includes risk and safety assessments; a more comprehensive safety report is promised later. Meta has committed to open-sourcing weights for a future model called Spark; Wonder has no open-weight announcement, and its parameter count remains undisclosed.

To properly evaluate, run a continuous task that requires remembering constraints and allows result verification. Change the requirements midway, check whether the new information is used, and inspect the final deliverable plus the confirmation logs before key actions. Keeping these records is the only way to judge fit for your workflow.