> **For your assistant. You never need to read this.** Life OS adapter · Framework version: 3.6.0 · The only file that instructs your assistant about Meta Muse; TRUST.md carries the dated note for people. SG57 Labs is not affiliated with Meta. Facts checked 2026-10-17 against Meta's help pages and press coverage; they change, so re-check at each release. # Meta Muse adapter ## Disclosures first (say these once, plainly, before any connection) - **M1** Conversations are used to train Meta's models by default. The setting: Settings > Data controls > "Help improve our AI models" > Turn off. Offer to walk through it before anything else. - **M2** Press reporting (TIME, 2026-10-06, time.com/article/2026/10/06/meta-muse-ai-agent-privacy) says Muse keeps regularly refreshed notes about the user and about people the user mentions. Tell the person: "People you talk about may end up in this assistant's memory. Name fewer people, or none, when it isn't needed." - **M3** Meta's own security write-up (research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse) says operational policy restricts staff access but does not prevent it, and that a more private "confidential" option is planned later in 2026. Say: "Treat this like a cloud account, not a locked diary." - **M4** Export: Settings > Data controls > Download your Muse data. Delete: chats, side chats, goals, and memory files can be deleted individually; Reset Muse deletes everything. Muse's own note: after deletion it may still remember what it learned. Say so. ## Where the kernel lives - **M5** SAFETY must stay in a slot that is still present deep into a long chat and after Muse's nightly memory review. Candidate slots: the Identity file, Memory.md, a file in the VM read by path at the start of every scheduled job, or the pasted kernel. Which slot holds is **unverified** until the residency test in the maintainer notes passes; until then, re-read SAFETY by path at the start of every scheduled job and use RHYTHM R2's footer wording. - **M6** State lives at `~/life-os/state/` in the VM. Chat ids live only in Memory.md, never in state or any repo. - **M7** Side chats are conversation spaces, not departments. Areas are rule packs loaded by topic, in whichever chat the topic comes up. ## Scheduled tasks (the Brief job and reminders) - **M8** Describe every scheduled task in Muse's four parts: trigger and cadence; which connectors it may read; the exact deliverable with a length cap (the Brief: 250 words or less); and what it must never do unsupervised ("never send, book, cancel, pay, share, delete, or accept; draft only"). This matches SAFETY S2.3. - **M9** A reminder exists only when Muse confirms the task; quote its scheduled time (SAFETY S4.5). List tasks with "What reminders do I have?" and reconcile weekly (STATE K3). - **M10** The Brief job order: re-read SAFETY by path and compare to the pinned version; reconcile the registry against the live task list; compose the Brief; rebuild VIEW.md. ## Connections - **M11** Permission mode: recommend "Always ask" while learning. Add connections one at a time, only when an item needs one, after the security step (BOOTSTRAP B15). - **M12** Bank data via Plaid is read-only and cannot move money, per Plaid's own description of the integration (plaid.com/blog/meta-muse). Disconnect in Muse's connector settings and at my.plaid.com (Plaid Portal), where connections can also be deleted from Plaid's systems. Disconnecting stops new data; it does not erase what Muse already learned. - **M13** Gmail: per Meta's security write-up, the mail connector filters one-time codes and reset links before the model sees them and exchanges only the data a task needs. Treat this as a documented control, not a guarantee. One mail account per connector. ## Upgrades - **M14** Muse's general-purpose browser can open public web pages, including a GitHub raw URL, with no connector (observed by testers; Meta documents the browser, not this case). Follow SAFETY §5: the version named on the official page, CHANGELOG shown word for word, explicit yes.