A folkfox special report.
Meta Muse AI memory:
the details that stay
Meta Muse AI memory and Spaces bring new ways to carry context and build a place to return to. We examine the company tests, privacy controls and what is actually available.
By the folkfox newsroom. Illustrations are conceptual.
The update in plain English
A useful assistant needs to remember that the meeting moved to Tuesday. It also needs to notice when Tuesday becomes Wednesday. On 6 November 2026, Meta said it was rolling out specialised versions of its existing models, trained using Muse Wonder, to the Meta Muse AI agent and Meta AI. The company calls the capabilities it is carrying across Active Memory and Active Focus. Meta AI Research
In a separate announcement, Meta introduced Spaces: prompt-built, persistent experiences inside the Muse app. Think of a trip planner, an order tracker or a small game that can be revisited and adjusted. Spaces are available where Muse is available, according to Meta. Sharing a space with other people was still being explored in the announcement. Meta newsroom
The model results are Meta's own evaluations; the company has not published enough detail in this announcement to tell us how reliably an individual user's tasks will go.
What is Meta Muse AI memory?
Meta describes Active Memory as the ability to carry relevant context from earlier conversations and recognise when it matters again. It says the updated models have been trained to distinguish temporary from lasting information, reason across time and identify details that have become stale. Meta AI Research
Imagine asking an assistant to organise a dinner after you have mentioned an allergy. A useful test is whether it applies that detail when suggesting food and revises its understanding if you correct an old preference. That is an illustration of the problem, not a demonstration we independently ran on Muse.
Memory can also become a liability when a guess hardens into a stored preference or an old fact keeps resurfacing. Meta says people can tell Muse to forget particular things, and its safety description says users can inspect, edit and download files in their virtual machine, including memory. The practical question is how easy those controls are to find and use in the version available to you. Muse launch and Meta safety explanation
The 90% memory result, properly labelled
For Meta Muse AI memory, the company reports that its current Muse model correctly used critical context in 90% of tasks on its Active Memory evaluation, compared with 76% for its previous Muse model. That is a 14 percentage point difference on the reported test. Meta also says the improved Muse models showed a 2.4-times reduction in forgetting information critical to a goal. Meta AI Research
Those figures are useful signals about what Meta chose to measure. They are not a promise that Muse remembers 90% of everything a person says, or that it will make the right choice nine times out of ten in ordinary life. The announcement does not give the sample size, task mix, uncertainty or enough detail to independently reproduce this particular evaluation.
The 90% and 76% scores concern the correct use of critical context. The 2.4-times statement concerns forgetting. The announcement does not explicitly say how the latter was calculated, so we do not reverse-engineer one from the other and present the result as Meta's method.
| Model described by Meta | Tasks correctly using critical context |
|---|---|
| Previous Muse model | 76% |
| Current Muse model | 90% |
Source: Meta AI Research. The announcement does not state the sample size or task mix. Difference: 14 percentage points.
What Active Focus adds
Active Focus is Meta's name for keeping an agent working towards a goal when it runs into a constraint or unexpected problem. The company says its improved Muse models showed a 67% improvement in completing everyday tasks where unexpected issues arise. Its release does not supply an absolute before-and-after completion rate for that statement. Meta AI Research
An improvement in completing disrupted tasks is a different measure from a 67% self-correction accuracy score. The first describes the outcome Meta says it measured; the second would suggest a specific success rate that the announcement does not establish.
Persistence needs a boundary, too. A useful agent might try a different train when a service is cancelled. If the alternative changes the price or the destination, a person may want a decision rather than an enthusiastic continuation. Meta says its safety system can allow, deny or request approval for actions according to the permissions and policies involved. People should examine the permissions they actually grant. Meta safety explanation
Muse Spaces: from chat to a place you can revisit
Meta Muse Spaces are custom experiences created by describing what you want in the Meta Muse AI app. Meta's examples include a music tracker, a wedding planner, a school dashboard and a small-business order tracker. A space can be revised through conversation and placed in the app's navigation. Meta newsroom
Meta says connected apps can help keep a space updated in the background. A delivery tracker, for example, is more useful if it can show a new status tomorrow. How current or complete its information is will depend on the relevant connection and should be checked before anyone relies on it.
The company says Spaces are available everywhere Muse is available. That is not a claim that Muse is available in every country. It also says it is exploring ways to share Spaces; the announcement does not establish that collaborative sharing has already launched. TechCrunch independently reported the launch and the build-by-description model.
A benchmark is a question, not a guarantee
Meta also reports a rise from 38% to 47% on its Personal Superintelligence Index for the previous and current Muse models, a difference of nine percentage points. The announcement does not explain the index's task composition or scoring in enough detail to compare it with a rival product's score. Meta AI Research
Academic researchers divide memory into several capabilities. LongMemEval tests information extraction, reasoning across sessions and time, knowledge updates and abstention. A system may retrieve an old detail yet fail to update it after a correction. A single headline score can hide that difference.
The authors of Memora designed a separate benchmark to examine remembering, reasoning and recommending as information changes. Their evaluation found repeated use of invalid memories in the systems they tested. It is evidence about those research setups, not a finding about Muse.
Mem2ActBench asks a related question: can an agent use long-term memory when choosing tools and their parameters? Its authors built a benchmark around simulated histories and tool-use tasks. It is a separate research benchmark, not an evaluation of this Muse rollout, and none of these papers verifies Meta's 90% result.
The questions worth asking of any memory claim are ordinary ones: what counted as a task, who selected it, what the model was allowed to see, and whether someone outside the company can repeat the test.
| Research | What it examines |
|---|---|
| LongMemEval | Retrieval, reasoning, updates and abstention across conversations. |
| Memora | Remembering, reasoning and recommending as information changes. |
| Mem2ActBench | Using memory to choose tools and their parameters. |
These are independent benchmarks, not tests of this Muse release.
What happens to the information Muse remembers?
Meta says a Muse user has a dedicated virtual machine and chooses which services to connect. Meta Muse AI memory is among the files people can inspect there, according to its safety explanation. That document says the ordinary Secure VM does not prevent Meta from accessing data when necessary to operate, support or secure the service. A separate Confidential VM, intended to prevent such access cryptographically, was described as planned and being tested with a small group. We found no current announcement establishing a general launch of that option. Meta safety explanation
The same explanation says conversations and VM data are not shared with Meta's advertising systems, while activity such as browsing a retailer can indirectly influence ads. It says sanitised interaction trajectories may be used to train models and that people can opt out in Muse settings. Those are Meta's descriptions of its design and policy; they do not amount to an independent audit of every data path.
There is a concrete reason to inspect sharing controls. In October, Ars Technica reported that researchers could discover thousands of user-created shared links through public DNS records. Meta told the publication it had paused that sharing feature while reviewing the report. The incident concerned discoverable links people had created, not evidence that every Muse conversation was public. The November Spaces announcement treats future sharing as something still being explored.
Meta Muse AI voice, avatars and availability
Meta announced voice conversations and expressive avatars for Muse separately on 5 November, with the page updated the following day. It says the feature is rolling out in the Muse app in the United States first, with options for voice alone or an on-screen avatar. Its announcement describes voice conversations on AI Glasses and Muse Charm as coming soon. Meta newsroom
The company has also discussed using Muse on mobile, the web and Mac. Availability differs by feature and place: the fact that a device can run a Muse experience does not establish that every new feature is already enabled on that device for every account.
Muse itself has expanded beyond its original US launch. For example, Meta announced availability in Singapore. Check the current product listing and settings where you live rather than assuming the US-first voice rollout describes the availability of Muse as a whole.
What a brand should test before connecting its work
For a small team, a persistent project space could bring a brief, deadlines and customer messages into one place. That convenience depends on access. Before connecting a work account, check who owns the account, whether the person connecting it has permission to share the material, and what the agent may read or do. Start with a low-stakes task and inspect the result.
Treat a remembered preference as a proposed instruction that can age. A campaign budget, a delivery address or a client's approval status can change between conversations. Keep the source of truth in the system your team already checks, and ask the agent to confirm consequential actions against the current record.
Teams using Meta Muse AI memory still need to check the data behind a dashboard. Our newsroom has examined how Muse might affect shopping discovery and what happens when an agent speaks in a marketplace. The same principle applies to a Space: test the outcome, not just the interface.
The questions to keep asking
For anyone trying Meta Muse AI memory, the questions are concrete: whether a remembered detail survives a long task, whether a correction takes effect, and when the agent asks before acting. The figures in Meta's announcement make a case for improvement on its own tests. They do not settle those questions for every user.
Before trying a connected workflow, review the apps and permissions involved, the saved memories, the training preference and the audit trail. Check whether a Space uses live data and what happens when a connection changes. For anything involving a customer, money or sensitive information, keep a person in the decision until the actual workflow has earned more trust.
For more context, read folkfox's Muse Wonder benchmark analysis, explore brand strategy and content marketing, or visit the newsroom.
Quick answers
What is Meta Muse AI memory?
Meta calls the capability Active Memory. It says the models can carry useful context from past conversations and recognise when earlier information matters to a task.
What does the 90% Muse memory score mean?
Meta says its current Muse model correctly used critical context in 90% of tasks on its Active Memory evaluation, compared with 76% for the previous model. The announcement does not disclose the task mix or sample size.
What is the 67% Active Focus claim?
Meta reports a 67% improvement in completing everyday tasks when unexpected issues arise. Its announcement does not provide an absolute completion rate for that result.
What are Meta Muse Spaces?
Spaces are custom experiences made by describing an idea inside Muse. Meta says they can be revisited and adjusted, and can use connected apps to keep information current.
Are Muse Spaces available everywhere?
Meta says Spaces are available wherever Muse is available. That does not establish that Muse is available in every country or that every feature has reached every account.
Is Muse voice available on AI Glasses?
Meta describes voice conversations on AI Glasses and Muse Charm as coming soon. Its Muse voice and avatar rollout began in the United States.
Can I control Meta Muse AI memory?
Meta says people can tell Muse to forget particular details and can inspect, edit and download VM files, including its memory. Check the controls and permissions in the version available to you.
Is Meta Muse AI memory private from Meta?
Meta says its standard Secure VM does not prevent the company accessing data when needed to operate, support or secure the service. A separate Confidential VM was described as planned in its safety explanation.
Does Meta use Muse conversations to train AI?
Meta says sanitised interaction trajectories can be used for model training and that users can opt out through Muse settings.
Work with folkfox
Make room for the human story
From emerging AI habits to a brand people remember, folkfox helps teams make a clear, considered next move.
Want folkfox in your Google results? Set folkfox as a preferred source.
Sources
- Meta AI Research: bringing specialised models to Muse and Meta AI
- Meta: introducing Spaces inside Muse
- Meta AI Research: introducing the Muse Wonder models
- Meta: voice and expressive avatars for Muse
- Meta: the original Muse launch
- Meta AI Research: how Muse safety was built
- TechCrunch: Meta introduces Spaces
- Ars Technica: discoverable links to shared Muse conversations
- Wu and colleagues: LongMemEval
- Uddin and colleagues: Memora memory benchmark
- Shen and colleagues: Mem2ActBench
- Meta: Muse availability in Singapore