WASHINGTON — A federal appeals court said Tuesday that users of long-term personal AI services must be allowed to carry more than old conversations and uploaded files when they switch providers.
Investors immediately calculated what that could cost Quarrylight Compute, whose consumer-agent franchise had been built partly on the assumption that relationships become harder to leave as the system learns more about a person.
By the closing bell, Quarrylight shares had fallen 16.4%, erasing about $1.97 trillion in market value. The calculation was based on the company’s roughly $12 trillion market capitalization at Monday’s close. Quarrylight remains one of the most valuable and profitable companies in the world.
The U.S. Court of Appeals for the District of Columbia Circuit upheld the central provision of a Federal Trade Commission rule requiring what regulators call “substantial functional continuity” when a customer moves a mature personal agent to a competing service.
The rule implements the Personal Agent Portability Act of 2040, which requires “effective switching” but left regulators and courts to decide what those words meant after an AI system had interpreted a person’s life for a decade or more.
The court held that transferring raw user records can be legally insufficient when the practical result is that the new agent must reconstruct years of user-specific context from the beginning.
Quarrylight said it would seek review by the full appeals court. The company argued that it had long allowed customers to export conversations, files, calendars, contacts, transaction records, explicit memories and user-entered preferences.
“The dispute isn’t over what users gave us. It’s over what our systems learned,” Quarrylight General Counsel Mara Ellison said outside the courthouse. She said the ruling risked turning proprietary intelligence generated by private models into regulated common infrastructure.
The court rejected the FTC’s broadest claims about access to internal representations. Providers will not have to transfer foundation-model weights, internal reasoning systems, raw proprietary embeddings or trade-secret architecture. Nor must a competing AI reproduce the old service’s personality, writing style, capabilities or behavior.
Instead, the rule requires a standardized continuity package describing user-specific state in a form another provider can use. That may include durable memories, current and revoked preferences, active goals, relationship context, delegated authorities, important exceptions to standing instructions and changes over time. The package must also distinguish what a user explicitly said from what the system inferred, and indicate uncertainty where it is material.
The difference is easier to understand through the experience of Elena Ruiz, a 54-year-old health-plan administrator in Richmond, Virginia, who tried to leave Quarrylight’s personal-agent service after using it for 13 years.
Ruiz exported more than 80,000 conversations, years of calendars and messages, and every explicit memory listed in her account. Her new service could retrieve the records, she said, but did not know that an old dietary preference no longer applied, that two names in her contact history referred to the same relative after a marriage, or which family disagreements had been resolved years earlier.
“I had all my data,” Ruiz said. “I still had to introduce myself.”
Consumer advocates call the missing layer “derived continuity state”: interpretations built from repeated interaction rather than facts typed into a profile. An agent may know which relationships remain important, which goals superseded older ones, which reminders help and which irritate, or when a general instruction has a long-standing exception.
None of those judgments is supernatural. Together, however, they can determine whether a system with access to someone’s records behaves like a service that knows the person’s present life or like a search engine pointed at an archive.
That distinction had become a major economic asset for Quarrylight. The company reported more than 600 million persistent consumer-agent relationships last year, and analysts estimated that more than nine in 10 premium customers whose agents had at least seven years of history renewed annually.
Investors called the advantage the “memory moat.” A customer was free to leave, but departure could mean spending months correcting obsolete assumptions and years rebuilding context. The older and more useful the relationship became, the greater the practical cost of switching.
Quarrylight earned subscription fees from premium agents as well as marketplace and service revenue from the commercial decisions those agents carried out. High retention supported pricing power, lower replacement costs and unusually high estimates of lifetime customer value.
Analysts said Tuesday’s ruling turns part of that private retention advantage into standardized consumer infrastructure: providers can still compete over models and service, but some accumulated understanding must now travel with the user.
“The market had treated memory as a compounding asset owned by the incumbent,” said Soren Vale, a technology analyst at Arc Meridian Research. “The court did not eliminate that asset. It changed how much of it can be used to keep the customer in place.”
Tuesday’s decline followed a separate 2.9% fall on Monday after a one-notch downgrade of Quarrylight’s senior credit rating. Traders said the court decision, not the rating action, accounted for the overwhelming majority of the second day’s move. Other large personal-agent providers fell between 2% and 5%, while smaller challengers and migration-service companies rose.
Basic portability had been standard for years. Through the late 2020s and early 2030s, regulators largely focused on whether people could retrieve chats, files, saved memories and permissions in machine-readable formats. Providers then built migration tools as persistent agents accumulated longer histories.
Those tools varied widely. A new service might receive every historical reference to a sibling without knowing whether the relationship was close, whether it had recently changed or whether an old conflict still mattered. Consumer complaints increasingly shifted from “I cannot get my information” to “I can take everything and still lose what the system understands.”
The 2040 law required effective switching across long-term personal-agent services. The FTC’s continuity rule, issued two years later, said a technically complete export did not satisfy the law if it left accumulated user-specific understanding functionally stranded inside the former provider.
Quarrylight and an industry group challenged the rule, arguing that some derived state could not be separated cleanly from the model architecture that produced it. The court said the agency could require an interoperable description of function without requiring the machinery itself.
Consumer lawyer Nia Calder compared the old system to changing doctors and receiving a box containing every historical lab result but none of the longitudinal record that explained which findings remained important.
Quarrylight called the analogy incomplete. “A medical chart is an expressible record,” Ellison said. “Some agent state exists only through proprietary computation. A provider cannot be ordered to export a process merely because the output feels personal.”
The court agreed that the two systems were not legally identical. Its opinion said regulators must judge whether a provider supplies usable continuity, not whether the receiving model arrives at identical conclusions.
“The statute protects a practical ability to switch,” the court wrote. “It does not promise an identical agent, compel disclosure of a model, or convert every machine-generated inference into user property.”
The decision also avoided deciding who owns an AI’s internal representation of a person. Users may not own that representation, the court said, but the statutory right to leave can still require enough portable state to make departure meaningful.
Implementation remains difficult because one person’s continuity includes information about other people. An agent’s understanding that a sibling is divorcing, a colleague is ill or a former partner should no longer receive updates may be useful to the customer while also revealing sensitive facts about someone else.
The FTC rule requires provenance tags, third-party data minimization and protected treatment for sensitive inferences. The appeals court directed the agency to provide more detail before the rule takes full effect, leaving providers to negotiate how much relational context can move without turning one person’s export into another person’s privacy breach.
The ruling also does not settle what Quarrylight may retain after a customer leaves. Fraud and safety records, aggregated learning and inferences distributed across other systems may be governed by separate rules. Consumer groups are expected to press for deletion rights, while providers argue that complete removal may be technically impossible or legally undesirable.
Quarrylight has 18 months to implement the final continuity standard unless further review delays it. Analysts said the company can still compete on model quality, reliability, tools, personality, security and the reach of its ecosystem.
What the ruling limits is a simpler advantage: requiring a customer who leaves to become a stranger again.
Reader forecast
Do you buy this future?
When do you think U.S. law will require long-term personal AI services to provide meaningful portability of accumulated memory and user-specific context across providers?
Relative to this report's 2044 date. One prediction per browser. Predictions cannot be changed after submission.