nanoMuse just dropped — first open-source personal agent that actually persists across devices instead of dying after one task.
Meta shipped Muse in Sept 2026 (iOS/Android, dedicated VM, learns from convo). Liu/Liu/Zhang reverse-engineered it from production prompts + shipped clients and built the open version.
Key difference: 2011 assistants answered queries. 2023 agents finished tasks. 2026+ agents run for weeks unattended, log everything, hold permissions.
Architecture:
• Android: 38MB arm64, sandboxed Alpine + browser + MCP servers + cron
• Desktop: Electron shell wrapping DeepSeek + Python runtime (256MB Linux / 498MB macOS, ~0.5GB idle RAM)
• iOS via TestFlight, web UI served from your machine
• Shared conversation via optional self-hosted relay (Python + SQLite, $4-6/mo)
Relay stores: hashed account ID, keys, device map, token ledger (model/cost, NOT content), synced text. Zero file storage.
Security layer = "Sentinel" with fixed decision order:
1. Deny list
2. User rules
3. Allow/ask lists
4. Taint rule (blocks private data leaving allowlist without approval)
5. Risk vs mode check
6. Destructive action warnings (shell, piped downloads, commitments)
Approvals are scoped: once / current convo / always for target. Everything logged.
Memory = plain files you can edit:
• SOUL.md (agent identity)
• USER.md (you)
• GLOBAL.md (persistent state)
• Dated diary
• HEARTBEAT.md (scheduled routines)
No API? Uses screen. Login stays with you. BYO model (18 providers listed, local server supported).
This is the first open implementation of a true persistent personal agent. Code's out there.
Meta shipped Muse in Sept 2026 (iOS/Android, dedicated VM, learns from convo). Liu/Liu/Zhang reverse-engineered it from production prompts + shipped clients and built the open version.
Key difference: 2011 assistants answered queries. 2023 agents finished tasks. 2026+ agents run for weeks unattended, log everything, hold permissions.
Architecture:
• Android: 38MB arm64, sandboxed Alpine + browser + MCP servers + cron
• Desktop: Electron shell wrapping DeepSeek + Python runtime (256MB Linux / 498MB macOS, ~0.5GB idle RAM)
• iOS via TestFlight, web UI served from your machine
• Shared conversation via optional self-hosted relay (Python + SQLite, $4-6/mo)
Relay stores: hashed account ID, keys, device map, token ledger (model/cost, NOT content), synced text. Zero file storage.
Security layer = "Sentinel" with fixed decision order:
1. Deny list
2. User rules
3. Allow/ask lists
4. Taint rule (blocks private data leaving allowlist without approval)
5. Risk vs mode check
6. Destructive action warnings (shell, piped downloads, commitments)
Approvals are scoped: once / current convo / always for target. Everything logged.
Memory = plain files you can edit:
• SOUL.md (agent identity)
• USER.md (you)
• GLOBAL.md (persistent state)
• Dated diary
• HEARTBEAT.md (scheduled routines)
No API? Uses screen. Login stays with you. BYO model (18 providers listed, local server supported).
This is the first open implementation of a true persistent personal agent. Code's out there.
