Meta’s Muse Wonder puts long-term memory at the center of its frontier AI push

Meta released Muse Wonder and Muse Wonder Full on October 12, bringing its Superintelligence Lab into the frontier-model competition after earlier releases focused on a lightweight model and an agent. The company says the new models match leading rivals on benchmarks and beat them on a test of long-term memory.

Coding scores help establish how Meta compares with OpenAI and Anthropic. Reliable recall could make its models useful for assistants working across weeks of changing requirements, preferences, and unfinished tasks.

Through the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. That puts it roughly alongside GPT-6 Sol and Opus 5.5, according to TechCrunch’s reporting. Developers will be paying premium-model rates, so the launch results need to hold up in production.

Benchmark parity needs workload testing

In its model card, Meta reports that Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra. The company also reports better performance on complex tasks, including coding, than its lightweight Muse Spark model.

Spark targets low latency; the larger Wonder models target more demanding work. That fits a familiar routing strategy: send simple classification or extraction requests to cheaper models and reserve expensive ones for debugging, planning, and difficult reasoning.

The launch details leave several operational questions unanswered. They don’t establish parameter counts, serving architecture, context-window limits, or measured latency. There’s also too little information to distinguish Wonder Full’s operational profile from Wonder’s. The quoted API pricing applies to Wonder; teams shouldn’t assume Full has identical economics.

A model can produce excellent answers while taking too long for an interactive coding tool. Competitive aggregate scores can also hide weaknesses in a particular language, repository structure, or tool-calling protocol.

For a coding assistant, give Wonder a real repository and ask it to fix a failing test. Measure whether the patch passes, whether it introduces regressions, how many tool calls it needs, and what the completed task costs. A fluent explanation of a bug has limited value if the patch still fails CI.

These are company-reported results. Independent testing will need to establish how consistently they carry over to practical workloads under comparable inference settings and budgets.

What Meta means by “active memory”

Meta puts unusual emphasis on what it calls active memory: the ability to retain and synthesize information supplied by a user over time.

“We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond,” research scientist Kavya Kopparapu says in a launch video.

The company reports that Muse Wonder scores above Fable, Opus, and Astra on MemoryBench, which evaluates continual learning and memory-construction frameworks. That’s relevant for agents. Many current applications can retrieve an old conversation but still struggle to determine which parts remain authoritative.

Consider an assistant helping a team migrate a service. On Monday, the team chooses PostgreSQL. On Thursday, a compliance requirement rules out the proposed hosting region. Two weeks later, the team changes its deployment strategy. Useful memory requires preserving the database decision, revising the infrastructure constraints, and discarding advice that depends on the abandoned plan. Finding a semantically similar passage won’t reliably handle all three.

A long context window can hold a large history, but repeatedly sending the full record adds cost and leaves the model to sort through stale information. Retrieval-augmented generation can narrow the material, though retrieval quality and document freshness remain separate problems. Persistent memory adds responsibility for deciding what to retain and how new information should change it.

The source material doesn’t specify Wonder’s memory implementation. Recall could depend on model behavior, an external store, summarization, retrieval, or a combination. “Active memory” also shouldn’t be read as evidence that the deployed model continually updates its weights for each user.

Developers need to know whether memory travels with an API integration, belongs to a particular agent product, or requires application-managed infrastructure. A strong benchmark result has limited practical value if a team can’t reproduce the setup.

Persistent memory needs controls

An assistant might retain an outdated customer preference, mistake a tentative suggestion for a final decision, or carry information from one project into another. A remembered error can affect many later answers without appearing in the user’s current prompt.

Enterprise deployments need provenance and lifecycle controls. Engineers should be able to identify where a retained fact came from, when it was updated, and who can access it. Users need a practical way to correct or delete it. Tenant isolation has to cover memory stores as well as ordinary application data.

Security testing should examine whether untrusted content can influence persistent memory. A malicious instruction in a retrieved document could potentially be stored as a durable preference or operating rule if the application doesn’t separate trusted user decisions from external material. This is a general risk for memory-enabled agents, not a demonstrated vulnerability in Wonder.

An evaluation should include contradictory updates, deletion requests, project boundaries, and deliberately planted false instructions. Recall accuracy alone won’t reveal whether an assistant can forget safely or revise its understanding correctly.

Meta’s memory result is promising because managing changing information is difficult. A higher MemoryBench score, however, doesn’t establish the access controls, retention policies, or deletion semantics an enterprise needs.

What premium pricing means for deployment

At Wonder’s announced rates, a request containing 100,000 input tokens and producing 10,000 output tokens costs $0.60: $0.40 for input and $0.20 for output. That excludes any separate tool, storage, or other charges.

The cost can add up when an agent repeats that pattern across multiple steps, especially if it keeps resending repository files, conversation history, and retrieved documents.

Memory could help if it lets an application supply a compact, accurate account of relevant prior decisions. That’s a plausible benefit, not an established property of this release. Memory construction and retrieval consume resources too, and an overly compressed summary can omit details needed to finish a task.

For most teams, it makes sense to start selectively. Keep low-risk extraction and routine transformations on a cheaper or faster model. Test Wonder where better reasoning or recall might reduce failed attempts, manual review, or repeated work.

Measure cost per successful task. A premium model can justify its price by completing a difficult job in fewer attempts, but a leaderboard won’t tell you whether it does that for your workload.

Where Wonder is worth testing

Wonder is Meta’s first frontier release since Alexandr Wang restructured its AI efforts in August 2025. The intervening releases, Muse Spark and the Muse agent, had a different product focus. Wonder gives Meta a high-end model offering to pair with its assistant ambitions.

Meta says it trained the models at its Prometheus data center in central Ohio, which came online earlier this year. The release draws on the company’s expanding compute capacity, though training infrastructure says little about API reliability, rate limits, or sustained inference performance.

The strongest reason to try Wonder is a workload where remembered decisions materially affect the answer: a long-running engineering project, a recurring analytical task, or an assistant handling evolving user constraints.

Give it a history with revisions and a decision that must be forgotten. Check whether it can explain which information guided its answer. Measure latency, task success, and the bill. Those results will tell a team whether Wonder’s memory advantage is worth paying for.