Two states put limits on teen chatbot memory as researchers try to improve it

What changed

A chatbot that remembers earlier conversations can be more useful than one that starts fresh every time. But when the user is a minor, should information from an earlier session still shape the next conversation? California and New York have arrived at different answers, according to an October 29 analysis by the Future of Privacy Forum.

The distinction matters because “memory” is not one action. A service can retain information, turn it into a summary, retrieve it later and use it to decide what to say. A rule about using an old disclosure is not necessarily a rule about deleting it. A rule that starts with memory switched off presents a different design choice again.

The same week, researchers published papers proposing more capable ways to organize and use agent memory—and examining what can go wrong when an agent attributes a remembered preference to the wrong person. Together, these developments sharpen a question for AI companies: not just can the system remember? but when should a memory be available, whose is it and what is it allowed to influence?

Two different limits for minors

As the Future of Privacy Forum describes it, New York’s S 9051B restricts the use of certain information obtained from minors—including information about health, well-being or personal matters—when it was acquired more than 12 hours earlier or in a prior session. That is a limit on the later use of specified information, not, on the evidence here, a blanket prohibition on every kind of chatbot memory.

California’s SB 1119 takes a different starting point in the FPF account: it restricts conversational memory by default, subject to stated exceptions and parental controls. The important words are “by default.” They describe the initial setting, not a claim that conversational memory can never be used.

These approaches would ask different things of a product team. A system responding to New York’s restriction needs to distinguish covered information from other details and recognize when it came from. One responding to California’s default needs to know whether conversational memory is permitted in the relevant circumstances before using it. Those are design implications, not a finding that any existing chatbot already handles either task well.

The policy question is also narrower than whether memory is good or bad for young users. A remembered preference might spare someone from repeating themselves; an old, sensitive disclosure might follow them into a conversation where it no longer belongs. The laws described by FPF make different judgments about where to put the initial constraint.

Research is making memory more selective—and more active

Three papers submitted on October 27 or 29 tackle another part of the problem: how an AI system should represent its past so that useful information can be found later. These are research proposals, not announcements that consumer assistants have adopted the methods.

ETAMem separates conversational memory into Event, Affect and Time views, then routes retrieval across them. Its authors report relative accuracy gains of 12.46% on LoCoMo and 7.96% on LongMemEval-S over their strongest baselines. The idea is that what happened, how it was experienced and when it happened need not be squeezed into a single way of searching. The reported gains are benchmark results; they do not establish how the system would behave with a particular person’s sensitive history.

ProactMem goes beyond answering a question about the past. It organizes history into recent events, temporal trajectories and behavioral regularities so an event-driven agent can identify opportunities to act. The proposed system checks trajectory claims against source events and replays history to assess whether a pattern fits the current situation. Its authors report an F1 score of 52.6% on ProactiveStream, a 21.4% relative improvement over the strongest of five external baselines.

That shift from answering to acting raises the stakes of an incorrect memory. If an assistant retrieves a mistaken detail in response to a question, a person may have a chance to spot it. If the assistant uses a supposed pattern to decide when to intervene, the mistake may shape its behavior before anyone asks about the underlying history. The paper’s source-checking approach addresses part of that reliability problem; it does not, by itself, decide whether the agent ought to use a particular memory.

A third paper combines episodic memories, which preserve events and context, with structured memories of entities and relationships. It also proposes consolidation to merge redundancies and remove outdated or conflicting entries. The authors report 90.2% accuracy on LoCoMo and 85.5% on LongMemEval-S, and say consolidation compressed the memory store by 16.34% without much loss of performance. Those are results for the authors’ method and tests, not a direct ranking against the other papers discussed here.

Consolidation addresses a practical issue: a system that keeps adding memories must eventually contend with repetition and conflict. But a shorter, cleaner store is not automatically a safer one. What gets merged, removed or preserved determines which version of someone’s history an assistant can later use.

A preference needs an owner

A separate paper submitted October 28 studies a “role grounding” attack involving failures to attribute a remembered preference to the right source or principal during social interaction. The problem is easy to picture: learning that one person prefers something does not mean another person does, or that the first person authorized the agent to act on their behalf.

This is a security concern as well as an accuracy concern. Memory can carry information across conversations; an attribution error can carry it across people. The research notes establish that the paper studies this attack, not how often it succeeds in deployed assistants. That distinction matters when assessing the risk.

It also complicates the policy challenge. Knowing when information was acquired is important to the New York restriction described by FPF. Knowing whose information it is—and whether a preference is being applied to the right person—is a separate requirement. A memory system could do one well and still fail at the other.

What it means

This week’s work points in two directions that have to coexist. Researchers are trying to make memory more structured, efficient and capable of informing an agent’s next move. Lawmakers are placing limits on when conversational history involving minors can be used. Neither direction cancels the other: better recall does not settle whether recall is appropriate, and a restriction on use does not eliminate the need to handle permitted memories accurately.

For readers evaluating an AI assistant, “has memory” is therefore an incomplete description. More useful questions are: What information can it draw from a previous session? Does it keep track of when and from whom that information came? Can it distinguish a record of an event from an inferred pattern? And what prevents an old or misattributed detail from steering a new response? The sources here show active work on pieces of that puzzle—not a finished solution.