Skip to content
AnalysisAI

Meta Tests Muse Wonder’s Memory With Facts That Change Midconversation

Meta’s new models lead its internal recall test. Muse Code’s documentation also explains what its coding assistant saves between sessions.

Alexandr Wang sits between two attendees and speaks into a microphone at Village Global’s Grove event.
Alexandr Wang, center, at Village Global’s Grove event on November 17, 2022. File photograph.Village Global / Wikimedia Commons · CC BY 2.0

When Meta tested Muse Wonder’s memory, its researchers changed facts partway through long conversations and asked models to recover answers. The test was published with Monday’s release of a new family of AI models designed to handle tasks across repeated conversations.

The company introduced Muse Wonder and Muse Wonder Full on October 12 for tasks that can require repeated conversations and many steps, including coding and handling documents. It calls their ability to retrieve relevant information across conversations Active Memory. Developers can add Wonder, currently offered in preview, to their own software or use it in Muse Code, Meta’s coding assistant. Wonder Full is available through selected Muse Code subscriptions.

On MemoryBench, Meta’s internal test, Wonder scored 85.6 and Wonder Full 85.5, compared with 81.5 for Anthropic’s Claude Opus 5.5 and 81.4 for OpenAI’s GPT-6 Astra. Those scores summarize judgments by another AI model using weighted grading rules. Meta’s description gives no question count or margin of uncertainty.

Each test conversation contains 64,000 to 100,000 tokens, the pieces of text a model processes. Researchers mix facts with unrelated conversation, introduce contradictions and revise details. Answering correctly requires tracking which information still applies. These are controlled conversations assembled for evaluation; the report does not describe a trial following customers over months of everyday use.

Research on AI memory has been examining that problem for several years. In LongMemEval, first published in October 2024, Di Wu and colleagues created 500 questions to test five abilities: extracting information, combining facts from separate sessions, understanding time, updating knowledge and declining to answer when evidence is missing.

Wu’s team used AI to draft conversations, then had people screen and edit them. Facts could emerge indirectly: a question about car insurance could reveal a recent car purchase. Researchers tracked where each fact appeared, giving them a record against which to check an assistant’s answer. Changing how conversations were divided and searched improved later answers. This older study did not test Wonder.

How Muse Code keeps project notes

Meta’s Muse Code documentation makes some of those choices visible. Its coding assistant keeps lasting project knowledge in ordinary text files. Meta suggests recording deployment procedures, unusual software behavior and “the reason a workaround exists.” Such notes give the assistant information a new session would otherwise lack.

At the start of a session, Muse Code loads an index and paths to as many as 48 other memory files. It reads individual files as needed. A background process can also supply a relevant note before the assistant responds. This arrangement lets developers preserve a short record of an earlier decision without feeding every past conversation into every new request.

Developers choose how widely to share those notes. Personal project memory is stored on their machine outside the shared code repository. Project memory travels with that repository to colleagues who copy it. A third option keeps personal notes across projects. These controls govern Muse Code; Meta’s consumer Muse service has its own application design.

Businesses building their own assistants also choose who retains conversation history. Meta’s Responses interface lets an application send earlier material with each request or ask Meta’s servers to store responses and reconnect them on later turns. The application supplies saved text when the model answers; adding a note does not require retraining the model.

Meta’s training-data policy distinguishes how customers pay. Wonder’s pay-as-you-go service uses Standard mode, which excludes customer content from model training. Muse Code subscriptions default to Contributor mode, which permits Meta to train future models on submitted prompts, code, files and responses. Subscribers can change that setting.

Meta bills developers for text they send to Wonder and text it returns. Its rates are $4 per million input tokens for Wonder and $20 per million output tokens. Reused input that qualifies for its prompt cache costs 20 cents per million tokens. The cache lets the service reuse processing of an earlier portion of a request, reducing the charge when an application sends that material again.

For people using Muse Code directly, Wonder Full is included in plans costing $50 or $200 a month, with different usage allowances. Meta says Wonder will reach its consumer Muse service soon, beginning with subscribers. Its October 12 announcement did not give a date for that rollout.

Common questions

What did Meta launch under the Muse Wonder name?

Meta introduced Muse Wonder and Wonder Full on October 12 for work that may span conversations and require multiple steps.

What does Meta mean by Active Memory?

Meta uses Active Memory to describe retrieving relevant information across conversations, such as a previous decision or project detail.

How did Wonder score on Meta’s MemoryBench test?

Meta reported 85.6 for Wonder and 85.5 for Wonder Full, with weighted grading by another AI model.

How long were the MemoryBench conversations?

Meta’s test conversations contained 64,000 to 100,000 tokens and included contradictions and updates to earlier details.

Did MemoryBench follow customers using Wonder for months?

No. Meta described controlled test conversations and did not report a months-long customer trial.

What does Muse Code store in memory files?

Meta’s documentation says Muse Code stores project knowledge in text files, which can include procedures and explanations for workarounds.

How many memory files can Muse Code reference?

Muse Code loads an index and paths to up to 48 other memory files, then reads individual files as needed.

What is the difference between personal and shared project memory?

Personal project memory stays on the user’s machine outside the repository. Shared project memory travels with that repository.

What training terms did Meta describe for pay-as-you-go Wonder?

Meta said pay-as-you-go Wonder uses Standard mode, which excludes customer content from training its models.

Did Meta give a date for the consumer Muse rollout?

No. Meta’s October 12 announcement said Wonder would reach consumer Muse soon, beginning with subscribers, without a date.

Sources

All news
Mark Zuckerberg in a dark sweater against a black background.
AI

Why Meta Reserves Thousands of GPUs for Its Biggest Training Runs

The Meta sign at the entrance to its Menlo Park headquarters, photographed in March 2022.
AI

Meta Seeks $35 Billion for AI Through Stock and Convertible Shares

Black-and-white portrait of Dave Wehner wearing glasses.
AI

Fireworks hires Meta strategy chief Dave Wehner as CFO

Have data to license or need a source?

Tell us what your company has or what your team needs.

Talk to SnowRock