An analysis of a 6.8 GB filesystem export from Meta's Muse AI agent has revealed the internal workings of the agent's Linux-based runtime environment. The export, obtained through an archive request, contains the root filesystem, internal documentation, integration code, and agent logs.
Analysis of Meta's Muse Runtime Reveals Internal Agent Structure and Experimental Features
The runtime environment, identified by the internal name "Hatch," includes a sandboxed Linux environment using bubblewrap. The agent's home directory contains several core files for managing identity and behavior, including SOUL.md, IDENTITY.md, USER.md, MEMORY.md, AGENTS.md, and TOOLS.md. Notably, the agent manages its own memory using plain Markdown files, which are organized into circumstances and experiences and made searchable via a PostgreSQL backend.
The filesystem also contains documentation for various integrations, including WhatsApp, Tailscale, and an experimental feature called "Meta Home Link." This experimental integration appears to utilize an ESP32-C5 chip with Wi-Fi and Bluetooth LE, suggesting potential capabilities for the agent to interact with devices on a local home network.
The export further revealed approximately 68 skill directories, which pair instruction files with command-line tools for tasks ranging from Google Workspace automation to media generation. Additionally, configuration files suggest Meta is developing connectors for services such as Slack, Dropbox, and Canva.
The researcher reported the findings, including concerns that sensitive runtime files and SSH keys could be exported through standard conversations, to Meta's bug bounty program. Meta categorized the report as "Not Applicable."
Sources
- How Meta's Muse works, revealed by the 6.8 GB filesystem it sent me （Hacker News Frontpage, 2026-09-22）