We Are Not Talking About AI Memory Enough

Our lives are shaped by what we remember. “Out of milk” turns into a quick note on the grocery list. “I miss my sister” spurs the long-overdue phone call. And, “remember my childhood nemesis?" prompts the funny anecdote recounted each holiday. Memories are foundational to how humans act: What is remembered informs what comes next.

Yet, in our interactions with personalized AI assistants, surprisingly little attention has gone to what these machines remember about us and how we can shape their memory to control what they do.

Recent innovations in memory systems powering AI assistants have dramatically expanded the set of actions AI assistants can take, including role-playing as long-term synthetic companions and iteratively executing complex, multi-step tasks over time. Companies are designing AI assistants that can store, retrieve, and update multidimensional information about users across conversations, while research on memory architectures––inspired by how humans remember events across time and space—accelerates in parallel.

These developments go beyond inferring user preferences as in social media-based ad targeting, opening the door to new forms of sustained, adaptive, and highly personalized interactions. Soon, an AI assistant may not only recall that I crave chocolate when stressed or have back-to-back meetings on Wednesdays, but may toss in that chocolate to my shopping list alongside a tailored joke it knows will make me chuckle: “Added a little mid-week treat. It’s not gourmet, but neither is your day.”

However, the more AI assistants can remember and do, the more is at stake when they fail. AI assistants have reportedly made unauthorized financial decisions, encouraged eating disorders, and distorted people’s sense of reality. Yet public discourse on curbing such failures rarely includes controlling assistant memory as a potential intervention.

As AI memory reshapes what assistants are capable of, and how users relate to them, we need to explore it as not just a technical feature but as a governance surface—a technological entry point that can double as a convening site for companies, policymakers, and civil society to enable greater user control over AI safety.

The real-world context AI assistants operate in is shaped by users’ lived experience—an area where users are, by definition, the primary experts. Giving users more control over what their assistants remember means they can test different memory configurations, repair or adjust their AI tools, and calibrate their own unique risk tolerance over time, reducing AI failures overall.

Reframing memory as a governance surface also aligns with developers’ goals to make personalization a core feature of AI assistants: the more personalized the assistance, the more likely people will find it useful.

Most AI assistants currently fall short of any such control. For instance, asking ChatGPT, “Based on all our past interactions, what do you know about my personal life?” yields an in-chat summary that doesn’t fully overlap with the “Saved memories” listed under the tool’s manual settings. This discrepancy may stem from ChatGPT’s “Reference chat history” feature that likely runs a sort of ad hoc search over all past conversations for in-chat queries. In contrast, “Saved memories” likely reflect a curated subset of user information stored for long-term reference.

On the front end, however, users are left guessing which memories really stick, how to change them, and whether doing so reliably affects how their assistant behaves. If users want ChatGPT to consistently remember a specific preference—say, a conversational style or boundary—should they state it directly in a chat session or manually edit memory settings? Turns out, neither option suffices.

While writing this article, I repeatedly instructed ChatGPT to ask me reflective questions one at a time instead of together in a list. Despite my in-chat instructions appearing as memory entries, each stating some version of “User prefers to be asked diagnostic questions one by one,” the assistant didn’t reliably follow through. One solution may be to add my preference in the “What traits should ChatGPT have” field. But that workaround only raises more questions about what ChatGPT counts as “memory” in the first place, and a creeping sense that I may not have a say in the matter.

Gemini’s rudimentary memory settings don’t fare much better, nor do Meta AI’s.

These assistants also offer no reliable way to prevent sensitive context about their user—say, private medical disclosures in one conversation—from carrying over to another about drafting a work email, potentially risking situations like unwantedly disclosing a disability to an employer. And if assistants remember something erroneously, the risks only compound.

At best, some tools allow users to turn memory use off for all chats, disabling the feature entirely. But what users need is an ability to segment memories and instruct the assistant: “Only remember health conversations when I explicitly ask for health advice, never for work-related tasks.”

User control over AI memory isn’t a silver bullet. Even a well-designed memory system can cause harm if it isn’t supported by complementary design safeguards and broader oversight. But treating memory as a governance surface opens up a new, concrete area for users to make personalized choices about a technology that is shaping their lives in unprecedented ways.

Unlike other interventions that are more opaque and sit further from user contexts—like changing AI model weights or training datasets—people can intuitively understand how memory shapes assistant behavior. They would no longer need to trace abstract risks through invisible data-sharing pipelines. Instead, they would be invited to simply notice what their AI assistant remembers, and decide if that’s okay. It’s an idea concrete enough to grasp, actionable enough to implement, and timely enough to prevent harms from spiraling.