nanoMuse: An Open-Source Personal Agent for Every Device You Own

nanoMuse. An Open-Source Personal Agent for Every Device You Own. By Guangyi Liu and two others, from Zhejiang University. Published on the sixth of October.

The report opens with a short history. Assistants from 2011 answered and then waited. Agents from 2023 did a task and then stopped. In September 2026, Meta's Muse showed something different: an agent for one person, with accounts, devices, memory, and a conversation that lasts. It was also closed, running in a vendor's cloud, in one country. An agent like that is expected to act on a person's accounts and devices, to remember them across weeks, to speak first when there is reason to, and to answer for what it did. The authors stress that this is a kind of software rather than a model, and that until now it had no open counterpart.

The report does three things in sequence. It defines the personal agent, in five questions and three horizons. It reads how Muse is built, working from Meta's public record and from a copy of its production prompt, with every statement marked by the source it came from. And then it presents nanoMuse, an open-source counterpart released under the GPL three point zero licence.

The design is one agent on every device a person owns, with hands on the phone's screen and on the computer's. Hands here means the agent operates those screens itself, rather than only talking about them. The devices are not separate assistants. They share a single conversation, carried over a relay that anyone can run, so the person is not dependent on one operator to keep it going.

Three further choices each answer one of the expectations set out earlier. Every action passes through a component called the Sentinel, which is where an agent that acts on someone's accounts can be made to answer for what it did. Memory is stored as files the person can read, so what the agent remembers across weeks is open to inspection. And the model underneath is the person's own choice, rather than one fixed by the system.

The evidence offered is descriptive rather than benchmarked. The size and cost of the system are given, explicitly as estimates. Several pieces are left open and laid out as a roadmap: memory that carries its provenance, an evaluation suite for the hands, and an open model to drive them. The abstract reports no measured results beyond that.

Read as a whole, the report is an argument about ownership as much as a system. The same category of agent that arrived closed and in one cloud is rebuilt here so that the relay, the memory and the model all sit where the person can see or replace them.