Two Paths to Personal AI
20 October 2026 | BunCosmos Agent Lab
—
Personal AI is beginning to move beyond the chatbot. Instead of opening a blank conversation, asking a question and leaving, the emerging idea is an agent that remembers the person it works with, keeps track of goals, uses tools, works in the background and develops continuity over time. Two projects with almost the same name make this shift unusually easy to see: Meta’s Muse, a commercial personal agent built as a complete product, and the independent open-source project that began as OpenMuse and now continues as nanoMuse.
The names are confusing enough to deserve clarification immediately. Meta’s Muse is the product: the personal agent people interact with. Muse Spark is a Meta model family introduced earlier in 2026 and used to power Meta AI and, at Muse’s launch, the agent itself. In October, Meta introduced the newer Muse Wonder and Muse Wonder Full models, with Muse Wonder expected to move into Muse progressively. OpenMuse, by contrast, was never Meta’s open-source release. It was an independent community project inspired by the product shape of Muse. That original repository is now archived, and its active continuation is called nanoMuse. [1][3][4][5][6]
Muse as a product
Meta introduced Muse in September 2026 as a personal AI agent designed to feel less like a tool that waits for individual prompts and more like a persistent helper that can carry work forward. People can talk to it in the Muse app or through WhatsApp, connect services, ask it to handle practical tasks and give it larger goals that unfold over time. Meta’s description is deliberately consumer-facing: the user should not need to understand agent frameworks, model routing or tool protocols in order to use it. [1]
Behind that simple interface is a fairly sophisticated architecture. Each Muse runs inside a dedicated Muse Secure VM, an isolated cloud computer containing the user’s workspace, browser, files and the credentials needed for connected services. The agent itself operates inside a restricted runtime rather than receiving unrestricted control of the entire virtual machine. Security-sensitive services remain outside that runtime, and credentials are kept away from the reasoning model. [2]
One of the most important pieces is Sentinel, a separate permission authority. Muse can propose an action, but Sentinel evaluates whether the action is allowed, denied or requires the user’s approval. The approval can be narrowly scoped to a single action, a task, a session, a period of time or a longer-standing permission. Meta also routes network access through this layer, so the agent’s ability to reach outside its environment is not supposed to depend only on the model behaving correctly. This is a useful design principle for personal agents: intelligence and authority are not the same thing. A model may be capable of doing something without automatically being permitted to do it. [2]
Meta’s research description of Muse goes beyond ordinary assistant behavior. The system can work in the background, use connectors, launch subagents and build tools for itself inside its controlled environment. Its memory is also meant to be inspectable by the user rather than existing only as an invisible personalization layer. Meta says users can inspect, edit and download files and Muse’s memory, while connected credentials are stored separately. [2]
The model layer is evolving as well. Muse Spark established Meta’s 2026 model line, while Muse Wonder and Muse Wonder Full were introduced in October as stronger models for agentic work. Meta describes Muse Wonder as adding Active Memory, which carries relevant information across conversations, and Active Focus, which helps the system keep pursuing a difficult goal, try alternative approaches, recover from errors and check in before important actions. Muse Wonder is expected to power Muse progressively rather than making Muse itself a different product. [3][4]
OpenMuse became nanoMuse
OpenMuse approached the same idea from almost the opposite direction. Instead of hiding the architecture behind a polished consumer product, it exposed the personal-agent stack so that people could inspect it, modify it and run it themselves. The archived OpenMuse repository described one agent with a name and a face that could live on a phone, remember its user, work on goals while the app was closed, use tools and ask before actions that were difficult to reverse. It supported OpenAI-compatible model endpoints rather than binding the agent to one model provider. [5]
That project has since moved to nanoMuse. The current project is an independent, GPL-licensed community effort and explicitly states that it is not affiliated with, endorsed by or derived from Meta or Meta’s Muse product. It now spans Android, desktop operating systems and a web interface, with the longer-term idea that the same agent can use different devices as different sets of “hands.” A phone can be one front door to the agent while a computer can provide a richer environment for files, shell commands, browsers or other tools. [6][8]
The important architectural idea is separation between the agent and the model that powers it. nanoMuse can talk to an OpenAI-compatible endpoint, so the reasoning model can come from a commercial API, a gateway or, where practical, a local model. That means the agent’s identity, memory, permissions, goals and tool setup can theoretically persist even if the underlying model changes. The “brain” is replaceable in a way that the agent’s accumulated identity does not necessarily have to be. [5][6]
nanoMuse also uses a Sentinel concept for tool permissions. In the open project, the agent can be configured to allow, ask or deny actions, while memory and audit information remain visible to the person running it. Its documentation describes goals that can continue over time, connectors for services such as mail and calendars, browser and shell capabilities, MCP support and skills. As with Meta’s design, the interesting safety idea is that the reasoning model is not supposed to be the final authority over sensitive actions. [5][6]
Self-hosting is optional rather than mandatory. The simplest mode lets the apps talk directly to a chosen model provider while the community relay handles sign-in and device discovery. A user who wants more control can run a private relay on a small server or spare computer, and a more technical user can run the full Python agent runtime in Docker. The project documentation says a modest home server or small VPS is enough for the relay because the heavy model computation normally happens at the model provider. [7]
Two philosophies of personal AI
The contrast between the two projects is more interesting than a simple “closed versus open” comparison. Muse treats personal AI as a finished service. Meta controls the model family, the secure cloud computer, the connectors, the permission system and the user experience. The benefit is that the difficult engineering is absorbed into the product. A person does not need to know what a container, API key, relay or MCP server is in order to ask the agent to do something useful.
nanoMuse exposes those pieces instead. The user can choose the model, inspect the code, decide where the relay runs, control the agent’s workspace and modify how the permission layer behaves. That brings more freedom and inspectability, but also more responsibility. Open source does not automatically mean private, secure or local. If the chosen model is a cloud API, the model request still leaves the device for that provider. If the user connects sensitive accounts, the safety of the whole system depends on how those credentials, permissions and machines are configured. Running your own agent means inheriting part of the security job that a commercial provider would otherwise handle.
This difference becomes especially important as personal agents gain longer memory and more authority. Meta can invest heavily in isolation, credential separation, prompt-injection defenses, approval interfaces and red-team programs at a scale an individual hobbyist cannot reproduce easily. An open system, however, gives technically curious users something commercial products rarely provide: the ability to inspect the machinery, change the model, move the agent to another machine and preserve the agent layer independently of one company’s product decisions. [2][6][7]
What the idea is really pointing toward
The deeper question is no longer simply which model gives the best answer. Personal agents suggest that an AI identity may eventually become a layer above models. A person could keep the same name, face, history, memory, permissions and long-running goals while the underlying reasoning engine is upgraded or replaced. The model would remain critically important, but it would become one component of a larger personal system rather than the whole identity.
This is why Muse and nanoMuse belong in an Agent Lab rather than only in a product-news section. Meta is demonstrating what happens when a company turns the personal-agent architecture into a polished service for ordinary users. nanoMuse demonstrates what the same broad idea looks like when the architecture is exposed to the user: model choice, self-hosting, readable memory, cross-device tools and a permission layer that can be modified rather than merely trusted.
For BunCosmos, the most interesting lesson is not that one approach should replace the other. A polished agent such as today’s commercial systems can already provide extraordinary continuity and capability, while open projects reveal how a future user-owned agent might be constructed underneath. Together they point toward a future in which “my AI” may mean more than a favorite model. It may mean a persistent agent identity that remembers, works, learns its role, uses different tools and can survive the arrival of the next generation of models without ceasing to be recognizably yours.
The central distinction: Muse delivers personal AI as a product. nanoMuse exposes personal AI as an architecture the user can inspect, modify and increasingly own.
Sources
Product capabilities and model availability can change. Sources below were checked for this edition on 20 October 2026.
1. Meta, “Introducing Muse: The World’s First Personal AI Agent Built for Everyone,” September 2026. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent
2. Meta AI Research, “How We Built Safety Into Muse,” September 2026. https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse
3. Meta, “Introducing Muse Spark,” April 2026. https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/
4. Meta, “Introducing Muse Wonder and Muse Wonder Full,” October 2026. https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
5. OpenMuse archived repository, OpenMuseAgent/OpenMuse. https://github.com/OpenMuseAgent/OpenMuse
6. nanoMuse current repository, nano-muse/nanoMuse. https://github.com/nano-muse/nanoMuse
7. nanoMuse self-hosting documentation. https://github.com/nano-muse/nanoMuse/blob/main/docs/self-hosting.md
8. nanoMuse “Every device” architecture documentation. https://github.com/nano-muse/nanoMuse/blob/main/docs/every-device.md