News · 2026-10-07

Muse’s memory report raises profiling questions, while Meta describes separate user environments

Reporting on Meta’s Muse describes persistent relationship files and instructions to infer users’ unstated goals, raising questions about how far an assistant’s memory should go. Meta’s public technical disclosure says each user has a separate cloud computer and that sanitized interaction records can train later models. The evidence supports a debate about profiling and pooled product learning, but does not show one user’s agent directly reading another user’s dossier.

Key facts

• TIME reports hourly updates to relationship files based on connected conversations, messages, and emails.
• Meta says users can inspect, edit, and download their agent’s files and opt out of model-training use in settings.
• Meta’s launch architecture retains operational access to user environments; a more confidential environment is future work.
• Primary source for Meta’s stated architecture and controls: its Muse technical disclosure.

A useful assistant needs some memory. Remembering a preferred meeting time or a recurring project makes future tasks easier. A file that infers interpersonal disputes, unspoken goals, and effective ways to persuade someone serves a broader purpose. The news is that those two forms of personalization are now being discussed inside the same agent product, with very different implications for consent.

In TIME’s report, available through AOL, reporter Harry Booth says he reviewed internal Muse instructions. The reported files describe relationships, shared interests, disputes, and “tensions and alliances.” Instructions also reportedly ask the agent to infer goals users have not explicitly stated and learn which prompts work best for them. These are attributed reporting findings about instructions, not an independent service-wide audit showing that every inference is accurate or every scheduled task runs as written.

The reported inclusion of people who do not themselves use Muse matters. If a user connects a mailbox, the assistant encounters correspondence about colleagues, family, and friends. Building a persistent relationship map from that material can turn an authorized connection to one person’s account into a source of inferences about many others. The concern does not require a data breach: it concerns the intended processing of ordinary authorized data.

Meta’s public account supplies a different, complementary layer of evidence. Each user has a dedicated virtual machine, a cloud computer containing files and agent state. Users can inspect, edit, and download that material, including memory. Meta says conversations, tool calls, and agent handoffs form interaction records used to train new checkpoints after sanitization removes key personally identifying information. Users can opt out through Muse settings.

Think of an assistant keeping its own notebook while the company studies sanitized work reports to improve the next assistant. That is different from letting every assistant open everybody else’s notebook. TIME reports instructions about shared lessons across instances; Meta says de-identified learning improves the product and is not directly shared between individual machines. Those statements leave questions about sanitization and persuasion, but they do not substantiate direct cross-user dossier access.

This distinction is particularly important because persistent agent memory and model training are separate systems. Editing a memory file changes material available to the current agent. Opting out of later training changes another processing path. Asking the system to forget a detail may change stored conclusions without deleting the original message. TIME reports that last behavior; Meta’s public technical post does not settle the complete deletion semantics. Readers should not assume that a single control performs all three operations.

Meta also says Muse conversations and virtual-machine data are not shared with its advertising systems. Its disclosure acknowledges an indirect route: an action on an outside website may affect ads from that site. It also says policies restrict employee access, while its current design does not prevent Meta from accessing machine data when needed to operate, support, or secure the service. The proposed confidential environment is a later-2026 plan, not a verified launch property.

The Hacker News discussion debates useful memory versus psychological analysis and nudging. One commenter claims relevant former Meta experience, but that identity is not independently verified. There is no representative poll or external audit in this discussion. The strongest defense of the product is that contextual memory makes an agent more helpful and users can inspect it. The strongest objection is that inspecting stored facts does not necessarily expose how inferred vulnerabilities or persuasive tactics influence future behavior.

Usage numbers should not substitute for evidence. The dossier contains conflicting descriptions of a four-million figure, while the linked gated Information headline refers to three million weekly users. Neither establishes how many dossiers were examined. The human briefing therefore leaves that count unresolved.

The existing Muse security story explains the separate boundaries around credentials and tools. Today’s privacy question extends that discussion: a technically isolated assistant can still make intrusive inferences. The test is whether people can understand, constrain, and meaningfully undo that processing, not simply whether their cloud computers are separate.

Primary source, verified: read the paper →

Ground Truth. (2026, October 7). Muse’s memory report raises profiling questions, while Meta describes separate user environments. Ground Truth. https://groundtruth.day/news/meta-muse-memory-report-raises-profiling-questions.html

@misc{groundtruth:meta-muse-memory-report-raises-profiling-questions,
  title  = {Muse’s memory report raises profiling questions, while Meta describes separate user environments},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/meta-muse-memory-report-raises-profiling-questions.html}
}

Topics: agents · privacy · memory · meta