Meta Muse: The Personal Agent That Keeps Working After You Close the App

Image: Muse at the center of a connected app ecosystem, guarded by Sentinel action controls.

Meta launched Muse on September 8 as a personal AI agent. It is for US adults 18 and over, via a dedicated iOS and Android app, muse.ai, and WhatsApp. AI glasses support is coming soon. Muse can keep working after you close the app: handling email, booking travel, making purchases, filling forms, negotiating, and pursuing long-term goals.

That autonomy changes the product question. A chatbot waits for a prompt; an agent needs context and a clear boundary around what it may do. Muse can connect to email, calendar, payments, shopping, health, and smart-home services, with permissions users can revoke.

A Personal Agent Needs a Personal Knowledge Layer

The important connection may not be a payment account or travel site. It is the layer that tells an agent what matters: preferences, commitments, recurring plans, and the background behind a goal. Without that context, “book my trip” is a form; with it, the request becomes a decision about timing, budget, and trade-offs.

That is the PKM angle. A BrainMap-style knowledge base can remain the durable record while an agent receives only task-specific context. Long-term goals should remain readable and revisable by their owner, not hidden in an agent’s history.

Secure Execution Still Needs Permission Design

Meta says Muse runs in a dedicated Secure VM. Its “Sentinel” controls require explicit authorization for sensitive actions such as sending email or spending money. The agent can prepare work autonomously, while a higher-risk boundary asks the user to authorize the final step.

A secure execution environment is not a substitute for judgment. Access to email, payments, health, or a smart home can be revoked, but it still deserves a narrow scope. Meta also offers an opt-out from training on interactions; an encrypted version is planned for late 2026.

Image: Mark Zuckerberg — Photo: Anthony Quintano, CC BY 2.0, via Wikimedia Commons

The Price of Delegation

The reported free tier allows roughly 100 million tokens per week. Power costs $20 per month, Maximum costs $100 per month, and a card is required even for the free tier. Muse reached No. 2 among US apps by September 10, according to TechCrunch.

More autonomy means more context and tool calls. A token allowance is not a guarantee of useful outcomes, so the right plan depends on which tasks are worth delegating.

A Practical PKM Rule: Keep Knowledge, Delegate Actions

Keep the source of truth outside the agent: decisions, constraints, preferences, and the current state of a project. Give Muse the smallest useful context, use revocable permissions, and require explicit authorization for sensitive actions. If a task can change money, identity, or a relationship, knowing the goal should never mean having unlimited authority.

Personal agents need personal knowledge, but ownership is the boundary. Let the agent act on a trustworthy, user-controlled memory that remains auditable and movable.

Sources: Meta, TechCrunch.

What do you think? Would you let an agent keep working after you close the app if its memory and permissions were both inspectable and revocable?