Meta's Muse Wonder Remembers Everything. Your Other AIs Don't.
Does Meta's Muse Wonder remember between conversations? Yes: its Active Memory carries context across chats, and it tops Meta's own MemoryBench.
AI Stella and Liam are AI hosts with AI-generated voices. About our AI hosts
Show notes
Does Meta’s Muse Wonder remember between conversations? Yes: its Active Memory carries context across chats, and it tops Meta’s own MemoryBench. In this episode of Agentic Pod, Liam and Stella break down the Muse Wonder launch, why no single model wins every task, and how Vilix AI gives Muse, Claude, ChatGPT and Codex one shared memory over MCP.
What we cover:
- Muse Wonder and Muse Wonder Full: what Meta launched, context window and API pricing
- Meta’s MemoryBench scores: Muse Wonder, Claude Opus 5.5, GPT-6 Astra
- What Active Memory and Active Focus actually do
- Who leads which benchmark row, and why the smart setup uses several models
- Why each lab’s memory stays inside its own app
- The Vilix AI fix: get_context, save_turn with source labels, newest memory wins
- User rules, project rules, projects and tasks, skills and the agent inbox
- Agencies: one Vilix AI account per client, so swapping in Muse Wonder keeps the client’s history
Chapters
Links & resources
Full transcript
Intro
Hey hey hey, welcome to Agentic Pod! I'm Liam.
And I'm Stella! And, you know, Meta just dropped Muse Wonder, and the big feature is memory. Like, it actually remembers you.
Finally, someone in tech who remembers my name.
Don't get attached. Okay, let's go.
The launch and the scoreboard
So what did Meta actually launch?
Two models from Meta Superintelligence Labs, Muse Wonder and Muse Wonder Full. Wonder is the fast one, it's out now through the Meta Model API and Muse Code, and Full is on Muse Code's higher plans. Both have a 1 million token context window, and the API is about 4 dollars per million input tokens and 20 per million output.
Okay, and I have the table, so let me read it, right? Meta's own MemoryBench. Muse Wonder, 85.6. Wonder Full, 85.5. Claude Opus 5.5, 81.5. GPT-6 Astra, 81.4.
Mhm.
So Meta made a memory test and Meta won the memory test.
It's their benchmark, yeah, independent numbers are still coming in. But honestly the bigger news is that memory is a benchmark now. A year ago it was a side feature, and now labs put it on the launch slide.
What Active Memory does
Okay so what does Active Memory, like, actually do?
It carries the important stuff across conversations and pulls it back in when it's relevant. So you mention your project on Monday, and on Thursday Muse just, uh, knows. And there's a second feature, Active Focus, which keeps the model going when a task gets hard, tries other approaches, and recovers from errors.
Honestly I need Active Focus for my Mondays.
Same, but they only ship it to models.
No model wins every row
So, so is Muse Wonder just the best now? Like, do I switch everything?
That's the fun part, because even Meta's own table says no. Wonder Full leads legal work on Harvey LAB, AutomationBench, DeepSWE and SWE-Marathon. Claude Opus 5.5 still leads Terminal-Bench, ProgramBench and JobBench. And GPT-6 Astra leads GDP.pdf, the document work.
Huh. So everybody wins something.
Right? Everybody gets a trophy. So the smart setup is using the best model for each job. You plan in ChatGPT on your phone, you build in Claude Code or Codex, and you hand legal and finance research to an agent running Muse Wonder.
Yeah, yeah. But wait, if I use three models, then...
Then you've got three memories.
One memory for every model
Right, because Muse remembers what I told Muse, and Claude remembers what I told Claude, and they don't, like, talk.
Exactly, each lab's memory lives in its own app. So every switch means re-briefing, the project, the decisions, the client's tone, what already failed. And the better each memory gets, the more you notice it doesn't follow you.
Okay so how do I fix that?
Vilix AI. It's one shared memory that every model reads over MCP. Before a reply, the tool calls get context, and Vilix AI hands back your relevant history, your rules and your open tasks. After the reply it calls save turn, and the exchange is saved with a source label, so you know it came from Muse or Claude or Codex. And when two tools disagree, the newest one wins.
Oh, nice. And, like, my rules?
User rules for your personal stuff, project rules for the stack and the tone, projects and tasks so every model sees what's open, and skills, so a procedure you build once gets followed by every agent. Plus the Vilix AI agent inbox, so you leave a note from your phone and your agents read it on their next run. And Muse is on the Vilix AI supported list too.
Agencies and recap
Okay, and agencies, you know, people building agents for clients?
That's the big one. You give each client their own Vilix AI account, and every agent you build for them, on any model, is wired to that client's key. So when Muse Wonder wins a task you run for that client, you swap the model, and the client's whole history comes along.
No re-briefing. I love it. Recap me.
Muse Wonder brings real memory to Meta's models, and no single model wins every task. So use the best one for each job, and connect them all to Vilix AI so your context, rules, tasks and skills follow you between Muse, Claude, ChatGPT and Codex. The article, the setup guide and the Vilix AI links are in the description. See you next time.
Bye!