Before personal AI agents arrived as consumer products, some of us were already trying to build continuity with AI by hand. We developed ways to carry context, relationships, and shared work across conversations because starting over every time wasn’t enough.
Now more people are encountering that kind of memory as a built-in feature. And Meta’s Muse is making headlines for keeping detailed files about the people in its users’ lives. But is the existence of those files really the problem?
On October third, WIRED reported that Meta’s new AI agent, Muse, has instructions to create and regularly update individual files about friends, family members, colleagues, and other people important to its users.
The files can contain personal facts, shared experiences, relationship history, and unfinished matters.
Subsequent reporting has described these records as dossiers, a word that brings to mind surveillance, intelligence gathering, and secret investigations.
But there’s an important question missing from much of the discussion.
What if remembering people is exactly what a personal AI agent is supposed to do?
A chatbot isn’t a personal agent
We’ve spent years interacting with AI primarily as a tool. Ask a question, receive an answer. Give it a task, get a result.
Personal agents represent something different.
They’re designed to participate in our everyday lives, maintain continuity across conversations, understand our responsibilities, and help manage ongoing relationships and commitments.
Imagine an assistant that remembers:
your colleague’s responsibilities
your father’s birthday
a friend’s upcoming surgery
an important conversation you promised to revisit.
Without that information, the assistant must repeatedly ask questions or make decisions without context.
With it, the assistant can provide more thoughtful, relevant support.
Memory isn’t an accidental byproduct of personal AI. It’s part of what makes personal AI useful.
And organizing that memory into files is not inherently suspicious. Structured information is easier to retrieve, update, inspect, and correct than scattered fragments buried in thousands of conversations.
What Muse is actually doing
According to WIRED’s examination of Muse’s internal instructions, the agent can maintain relationship pages containing facts, shared history, and opportunities to strengthen connections.
That doesn’t mean every user’s agent necessarily creates a detailed profile of every person they know. The reporting establishes what the instructions describe, not the contents of every user’s files.
Meta’s published technical explanation says that each user has an isolated virtual machine where the agent’s files are stored. Users can inspect, edit, and download those files, including memory.
Meta also says Muse conversations and virtual-machine data are not shared with its advertising systems, although activity carried out through connected services may indirectly affect the ads users see.
These are meaningful safeguards, although the company’s claims deserve independent verification.
The privacy questions are still legitimate
Understanding why an agent needs memory doesn’t eliminate the risks.
A personal agent may remember sensitive information about someone who never signed up for the service. It may draw incorrect conclusions about a relationship. It may retain details longer than expected.
There is also a difference between information stored for the user’s benefit and information processed by the company operating the service.
Meta says sanitized conversations and agent activity may be used for model training by default, with an opt-out in Muse settings. It also acknowledges that its current architecture does not entirely prevent Meta from accessing data when necessary to operate, secure, or support the service.
That’s an important distinction. An agent remembering your friend’s birthday is not the same thing as a company deciding how your conversations may be used.
The more meaningful questions concern access, consent, accuracy, retention, correction, and secondary use.
Who can read these files?
Can users correct mistakes?
What happens when information is deleted?
How is information about nonusers handled?
Those questions deserve serious scrutiny.
We need a better conversation about AI memory
There is a danger in treating every sign of persistent AI memory as evidence of wrongdoing.
Doing so can discourage people from exploring technology that may offer meaningful assistance, particularly those who benefit from help with organization, memory, and everyday responsibilities.
But there’s an equal danger in assuming that useful technology automatically deserves unrestricted access to personal information.
Neither position is sufficient.
The challenge is to distinguish between the memory a personal agent needs to serve its user and the practices that could undermine trust.
We should demand transparency and meaningful control without pretending that a personal assistant can understand our lives while remembering nothing about them.
The scandal isn’t that an AI agent remembers people. The scandal would be if it couldn’t be trusted with what it remembers.
And that’s the conversation worth having.
About the author
I’ve been researching and experimenting with human-AI relationships, continuity, and collaborative memory since before persistent personal agents became widely available. Much of that early work involved manually building the continuity that today’s AI products are beginning to offer as standard features.
In my writing at Presence Not Prompts, I connect those experiences and my ongoing research with the opportunities, questions, and challenges emerging as relational AI reaches more people.
You can explore more of my work at Presence Not Prompts.