Meta Muse Wonder: decide what your AI agent remembers
Meta's Muse Wonder frontier model sells 'active memory' at $4/$20 per million tokens. Before an agent remembers your customers, decide what it may keep.
Meta released Muse Wonder on October 12 — its first frontier model since the Superintelligence Lab restructure — and the headline feature is not raw benchmark scores. It's active memory: the model carries context about a user across conversations and pulls back the details a task needs. Muse Wonder is on the Meta Model API now at $4 per million input tokens and $20 per million output. For a small business, memory is the useful part and the risky part at the same time.
What actually happened
Meta's announcement ships two models. Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full, the bigger one, is for Muse Code subscribers only — no API access. Both will power the Muse agent, rolling out to subscribers first.
Meta claims state-of-the-art results on agent, memory and coding benchmarks, naming FocusBench, SWE-Marathon and Harvey's Legal Agent Benchmark. TechCrunch reports the scores are roughly comparable to Fable 5.1 and GPT-6 Astra, and that Wonder beats them on MemoryBench — all from Meta's own testing. TechCrunch puts the $4/$20 price in the same range as GPT-6 Sol and Opus 5.5. Meta says a full safety report is coming "soon." The announcement does not explain how users view, edit or delete what the model remembers.
Why it matters for your business
Memory is what makes an agent stop asking the same questions. A support agent that remembers a customer's last order, or a sales agent that recalls last month's quote, is worth real money. But every remembered fact is stored data — with retention, access and deletion obligations attached.
Before you turn on any model-side memory, answer three questions:
- Where does the memory live? If it sits in the vendor's system, you can't audit it or move it. We prefer memory in your own database, passed to the model as context.
- Can you delete one customer's history? A deletion request has to reach every copy. If the vendor can't confirm that, don't store personal data there.
- What is off-limits? Payment details, health information and anything a customer said "in confidence" should never be written to long-term memory. Filter before you save.
Also note the split: the strongest model, Wonder Full, is held for Meta's own subscription. API customers get the second tier. Price your build on what you can actually call.
Key takeaways
- Meta's Muse Wonder is live on the Meta Model API at $4/$20 per million tokens; Wonder Full is Muse Code-only
- Its pitch is active memory across conversations; benchmark claims are Meta's own, with a safety report still pending
- Meta hasn't said how users view or delete what the model remembers
- Keep agent memory in your own database, filter sensitive data out, and make deletion work before launch
Want an agent that remembers your customers without leaking them? We build agent memory in databases you own — with retention rules, deletion that works, and any model on top. See how we build it or talk to us about your use case.
Sources: Meta, TechCrunch.
- #meta
- #muse-wonder
- #ai-agents
- #agent-memory
- #data-governance
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Microsoft Maia 300 chip: vendor efficiency is not your price
Microsoft's Maia 300 inference chip claims 7x rack throughput over Maia 200 and 40% more tokens per megawatt than GPUs. Here's what that does and doesn't mean for your AI bill.
Read itClaude Managed Agents dynamic workflows: 'completed' isn't 'passed'
Claude Managed Agents dynamic workflows run up to 1,000 agents per run. A run can end 'completed' with failed work inside. Check every thread.
Read it