Forget Me: How Muse Handles Deletion Requests
Muse lets users tell it to forget specific things it has learned. That's a simple-sounding feature that raises real questions about what memory in an agent actually is.
An agent that's supposed to remember your preferences across sessions — what Meta describes as one of Muse's core capabilities — creates an obligation most chat products never had to solve seriously: what happens when a user wants a specific piece of that memory gone? Meta's stated answer is direct. Users "can always tell it to 'forget' specific things it's learned." No settings menu buried three screens deep, no support ticket — a conversational instruction that's supposed to work the same way any other request to Muse works.
Why this is harder than it sounds
For a traditional app, "delete my data" usually means locating a database row and removing it — a mechanical, well-understood operation. For an AI agent whose memory of you is built from accumulated context across conversations, tasks, and inferred preferences, "forget this specific thing" is a much fuzzier operation. If Muse learned that you prefer aisle seats from a flight-booking conversation three weeks ago, and you tell it to forget your seating preference, does that memory get cleanly excised, or does it linger in some derived form — a general sense that you "care about seat comfort," inferred and retained even after the specific fact was supposedly deleted?
This is the same category of problem that has dogged large language model unlearning research more broadly: it's straightforward to delete a row in a database, and it's considerably harder to guarantee that a learned system has genuinely stopped "knowing" something it was trained or conditioned on, rather than just stopped saying it out loud when asked directly. Meta's announcement doesn't detail the technical mechanism behind "forget" requests — whether it's a hard deletion from a structured memory store, a soft filter that suppresses the information without removing it, or something else entirely. That's a real gap in what's publicly known, and it's worth being explicit that it's unresolved rather than assuming the feature works exactly as the one-line description implies.
How it fits with Muse's broader privacy stance
"Forget me" sits alongside two other stated privacy commitments in Meta's launch materials: users can opt out of having their interactions used to train Meta's models, and Muse doesn't share conversations or VM data with Meta's ad systems, a boundary we cover in the context of Muse's small-business connectors in Ad Data and Agents. Taken together, these three commitments — no ad-data sharing, opt-out of training, and on-demand forgetting — form Meta's answer to the most obvious objection a Meta product raises for a privacy-conscious user: that this is the same company whose entire consumer business model has historically run on retaining and monetizing behavioral data.
A "forget me" feature is as much a trust signal as it is a technical feature. Meta building it as a conversational, low-friction request rather than a buried settings option suggests the company understands that the credibility of its other privacy claims rests partly on how easy this one is to actually use.
The strongest version of this commitment — one that would make deletion cryptographically final rather than policy-dependent — is tied to Muse Confidential VM, the end-to-end encrypted environment Meta has said is coming later in 2026, which we cover in Muse Confidential VM: What End-to-End Encryption Will Change. Until that ships, "forget" requests operate within Meta's existing infrastructure and policy commitments — real, but dependent on Meta continuing to honor them, rather than being technically unable to do otherwise.
The open question
What Meta hasn't published is any verification mechanism — a way for a user to confirm that a "forget" request actually took effect, rather than trusting the confirmation message Muse gives after the request. For a feature this central to the product's privacy pitch, an independently auditable way to verify deletion actually happened would be a meaningful next step, and it's not something the launch coverage — including TechCrunch's roundup of everything new coming to Muse — has addressed. That's a reasonable thing to keep watching for as the product matures past its first month.
How this compares to deletion rights elsewhere
Regulatory frameworks like Europe's GDPR have spent years establishing what a legally meaningful "right to erasure" actually requires — verifiable deletion, defined timelines, and in many cases a documented process a company can be held to. Muse's "forget" feature, as described in launch materials, is framed as a conversational convenience rather than a formal compliance mechanism, which raises a real question about how it interacts with regions where users have a separate, legally enforceable deletion right that exists independent of whatever Muse's own memory architecture happens to support. Meta operates in the EU and has extensive experience navigating GDPR requirements across its other products, so it would be surprising if there weren't a more formal deletion pathway operating underneath the conversational "forget" request for users in regulated markets — but that underlying mechanism, if it exists as something distinct from the consumer-facing feature, hasn't been detailed in the sources reviewed for this piece.
There's also a practical usability angle worth naming. A feature that only works when a user remembers to explicitly invoke it puts the burden of privacy management on the user's own vigilance — you have to remember that Muse learned something you'd rather it hadn't, and you have to think to ask it to forget that specific thing. That's a meaningfully different posture from an interface that periodically surfaces what it has stored and lets a user review and prune it proactively, the way a browser's saved-password manager or a smart-speaker's voice-history page does. Whether Muse eventually adds that kind of proactive memory-review surface, rather than relying solely on users to initiate "forget" requests, is a reasonable feature request for a product this early in its life, and one worth watching for in future updates.
Part 9 of 30 in the Meta Muse series. Previously: Muse Confidential VM: What End-to-End Encryption Will Change. Next: Muse Realtime Avatar: Giving a Voice Agent a Face. Browse more builder insights or explore AI skills at aiskill.market.