This paper defines the notion of a personal agent through five guiding questions:
- Who does the agent serve?
- What capabilities does it provide?
- How does it retain memory?
- What is its interaction modality?
- Where is it deployed? It then outlines three development horizons: a short‑term usable prototype, a mid‑term unified cross‑device memory layer, and a long‑term autonomous, auditable digital assistant.
Meta’s Muse, announced in September 2026, runs as a single cloud instance tied to a user’s accounts and devices, with all data stored in the vendor’s cloud, limiting transparency and portability.
nanoMuse is the open‑source counterpart released under GPL‑3.0, featuring:
- Multi‑device deployment: a lightweight agent runs on every personal device (phone, laptop, etc.).
- Unified conversation: instances communicate through a self‑hosted Relay server, yielding a single cross‑device dialogue stream.
- Sentinel guard: every external action (sending email, launching an app) must pass through a Sentinel, providing a security audit trail.
- File‑based memory: memories are stored as ordinary files that the user can read, back up, or delete, ensuring provenance.
- Pluggable model: the inference engine is not fixed; users may plug in any open‑source LLM (e.g., Llama, Mistral) to meet privacy and cost constraints.
Size and cost estimates suggest a 7‑billion‑parameter model, with inference costing roughly $0.0005 per thousand tokens, runnable on consumer‑grade hardware.
An evaluation suite is provided to benchmark the agent’s screen interaction, file manipulation, and dialogue coherence, followed by a roadmap that includes strengthening memory provenance, standardising the Sentinel API, and fostering a community‑driven model repository.
Review