What Is Active Memory? Muse Wonder's Long-Term Memory Explained
Active Memory is Muse Wonder's long-term memory: it retains and surfaces important context about you across conversations spanning days, weeks, and months.
By Professor Muse9 min read
Active Memory is the long-term memory of Muse Wonder, Meta’s frontier model announced on October 12, 2026. It lets the model retain and synthesize important context about you across conversations spanning days, weeks, and months, and surface it when a later task needs it. Instead of starting every chat from zero, Wonder builds continuity with you over time.
Key takeaways
- Active Memory is one of two headline capabilities of Muse Wonder, alongside Active Focus. Meta says both models in the Wonder family, Muse Wonder and the larger Muse Wonder Full, come with it.
- It is Meta’s name for the model’s ability to retain and synthesize information you share across conversations spanning days, weeks, and months, and to pull the relevant pieces back up when a later task needs them.
- On MemoryBench, a benchmark for continual learning and memory construction, Muse Wonder scores higher than Claude Fable, Claude Opus, and GPT-6 Astra, according to Meta’s own testing.
- Meta has not publicly explained the internal mechanics of Active Memory, what exactly gets stored, or the full set of user controls, so treat specific claims about how it works as unverified until Meta publishes more.
- Active Memory is only available through Muse Wonder, which requires a paid Muse subscription ($20/month Power or $100/month Maximum). The free tier runs Muse Spark 1.3.
What Active Memory actually is
When Meta announced its first frontier models, Muse Wonder and Muse Wonder Full (described by Mark Zuckerberg in an X thread as bringing Meta “to the frontier”), Meta put two capabilities at the center of the launch: Active Memory and Active Focus.
Active Memory is Meta’s term for the models’ ability to retain and synthesize information you give them over long periods. In Meta’s own framing, reported by TechCrunch, the models can keep track of what matters to you across days, weeks, months, and beyond, and bring the relevant details back at the right moment, when a new task needs them.
Research scientist Kavya Kopparapu put it this way in a video accompanying the launch: “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.”
Think of it less like a filing cabinet and more like a very attentive colleague. You tell them your preferences, your ongoing projects, and your constraints over months of working together, and they don’t need you to repeat the setup every Monday morning.
How it’s different from a context window
This is where most people get tripped up, so let’s make the distinction crisp.
A context window is a model’s short-term working memory within a single conversation. Both Muse Spark 1.3 and Muse Wonder have a context window of roughly 1 million tokens, meaning that inside one chat, the model can consider an enormous amount of text at once. But when that chat ends, that working memory goes with it. The next conversation starts from a blank slate.
Active Memory operates across conversations. It is the persistent layer: the model keeps important context from past sessions and retrieves it when it’s relevant to what you’re doing now. Meta says Wonder scores higher on MemoryBench, a benchmark designed to test continual learning and memory construction frameworks, than competing models including Claude Fable, Claude Opus, and GPT-6 Astra.
One caveat: those benchmark results come from Meta’s own testing. Independent verification will take time.
How it compares to ChatGPT’s memory
If you’ve used ChatGPT’s memory feature, Active Memory will sound familiar. And it should. OpenAI’s system also carries details across conversations, through a combination of saved memories you explicitly ask it to keep and insights it extracts from chat history on its own. In January 2026, OpenAI upgraded it to surface things you discussed up to a year ago.
What’s different here is mostly positioning and architecture. ChatGPT’s memory is a product feature layered onto the assistant; Meta is presenting Active Memory as a core capability of the Wonder models themselves, trained for long-horizon tasks and measured on a purpose-built memory benchmark. Whether that translates into a meaningfully better experience in day-to-day use is something reviewers and users will have to test now that the models are out.
What Meta has not said matters too: the company has not published the internal mechanics of Active Memory (how memories are stored, weighted, updated, or retrieved inside the system), so be skeptical of any third-party technical claims about how it works under the hood.
What Active Memory enables
The point of remembering, in Meta’s telling, is agency. Wonder is meant to power the Muse agent through long-running, real-world tasks where losing context mid-way would be fatal. Meta’s examples include navigating an insurance claim, understanding a lease, and making sense of someone’s finances. These are jobs where the relevant information arrives in pieces over days, and an assistant that forgets each session would make you redo the setup every time.
Imagine telling Muse about a project in March, your budget, your constraints, your collaborators, then coming back in June and having it pick up with all of that already in place. Or asking it to compare a new contract against the lease you walked through with it two months earlier, without re-uploading anything. That’s the promise: continuity turns the assistant from a tool you instruct into a partner that accumulates knowledge.
That continuity is also why Meta pairs Active Memory with Active Focus: memory provides the context, and focus provides the persistence to keep working through hard tasks. We break down the other half of the pair in our Active Focus explainer.
Privacy and controls: what’s known
Long-term memory is useful, and it raises fair questions. Here’s what Meta has said, and what remains unknown.
What you can check and control
According to Meta’s help center, Muse stores things you share in a “Memories” file so it can personalize responses. You can view it by tapping the Assistant icon, then Identity, then Memory. You can edit or delete entries directly in the Memory file. There’s also a full “Reset Muse” option, which permanently deletes chat history, files, and active tasks.
One important caveat
Meta’s own help page warns that after you delete something, Muse “may still remember information it learned from what you deleted.” Editing the file removes the stored text, but the model may have already absorbed what it learned into its working understanding of you, so don’t treat deletion as a complete eraser.
Memory is not the same as training
Meta also offers a “Help improve our AI models” toggle in Muse’s data controls, which controls whether your chats are used to improve the company’s AI models. That’s a separate question from memory. It’s the other half of the data conversation, and worth checking in your settings either way.
What hasn’t been disclosed
Meta has not detailed exactly what qualifies as “important” enough to be remembered, how long stored context persists, whether memory works the same across every surface where Muse appears, or the full mechanics of how stored context is retrieved and applied. Until that documentation exists, the safest assumption is the simple one: an assistant that remembers is an assistant worth being thoughtful about. Keep sensitive information out of conversations unless you have a reason to share it: credentials, financial details, other people’s private information.
Muse’s memory capabilities have also drawn scrutiny: reports have described incidents involving the assistant’s handling of private information, which Meta disputes while saying it has strengthened safeguards. Long memory amplifies both the value and the stakes of those safeguards.
The bottom line
Active Memory is Meta’s bet: a better memory moves assistants forward more than a bigger brain. The idea is straightforward: a model that remembers what matters to you across months can handle the kind of long, messy, multi-session tasks that today’s chatbots fumble.
Whether Meta’s implementation lives up to the billing will depend on details the company hasn’t shared yet: exactly what’s stored, how reliably it’s retrieved, and how much control you have over it. But if you’re wondering how Wonder compares to the free Muse Spark 1.3 model, Active Memory is arguably the biggest reason to care: it’s the clearest answer to what the $20/month Power subscription buys you beyond raw capability.
Frequently asked questions
Is Active Memory available on Muse’s free tier? No. Active Memory is a capability of the Muse Wonder models, which are available in the Muse app through the $20/month Power or $100/month Maximum subscriptions. The free tier runs Muse Spark 1.3.
Do both Muse Wonder models have Active Memory? Yes. According to Meta’s announcement, both Muse Wonder and the larger Muse Wonder Full are equipped with Active Memory (and Active Focus). Wonder Full is currently limited to Muse Code subscribers.
Does Active Memory mean Meta is training on my conversations? Not automatically. Memory (keeping context to personalize your experience) is separate from model training. Meta provides a “Help improve our AI models” toggle in Muse’s data controls that controls whether your chats are used for training. They’re two different things, and both are worth checking in your settings.
Can I see or delete what Muse remembers about me? Yes, per Meta’s help center: tap the Assistant icon, then Identity, then Memory to view the file, and you can edit or delete entries directly. One caveat from Meta’s own documentation: the assistant may still remember information it learned from what you deleted. A full “Reset Muse” option permanently deletes chat history, files, and active tasks.
How is Active Memory different from a big context window? A context window is short-term memory within a single conversation. Muse Wonder’s is roughly 1 million tokens. Active Memory is the persistent layer across conversations: it retains and retrieves important context over days, weeks, and months after individual chats have ended.
Does “active” mean it’s always recording? Meta hasn’t published the internal mechanics, so nobody outside the company can say precisely how memories get selected, stored, or retrieved. What Meta has said is that the model focuses on remembering “what’s important to you” and surfaces it when a later task needs it.
How does Active Memory compare to ChatGPT’s memory? The goals are similar: both carry context across conversations. ChatGPT uses saved memories plus automatically extracted insights from chat history. Meta’s differentiator is that Active Memory is a core model capability benchmarked on MemoryBench, where Wonder outscores competitors. Those are Meta’s own test results, though, and independent verification is still pending.
