highbeamSubscribe

Policy brief 10 · Chatbots

Personal Memory Control

A chatbot's memory of you should answer to you.

All policy briefs

Americans now bring chatbots their health questions and relationship troubles, and the leading chatbots remember what they are told. Users cannot reliably see that memory, take it elsewhere, or keep it out of advertising. Congress should make memory visible, let users export the full record, make deletion real, and require separate consent before conversations feed ads, sale, or training.

The Problem

About a third of American adults turned to an AI chatbot for health information or advice in the past year, and 13% of all adults have uploaded personal medical information to one, though 77% worry about its privacy.1 OpenAI's chief executive has said people use ChatGPT "as a therapist, a life coach," and that no legal privilege protects those talks as it would with a doctor: "we haven't figured that out yet."2

Memory makes these tools more useful, and it is becoming the default. In April 2025, ChatGPT began drawing on all of a user's past conversations; that August, Google gave Gemini the same ability and switched it on by default.34 A passing fear or a hard month can harden into a lasting profile that shapes every later answer.

Those records now feed advertising. Meta announced that from December 16, 2025, it would use conversations with its AI to personalize ads, excluding sensitive topics such as health.5 It offered no opt-out and exempted only the European Union, Britain, and South Korea, which have stricter privacy laws.6 An American's chat with Meta AI could shape their ads; a Parisian's could not. Three failures leave users without control:

  1. The memory stays hidden. OpenAI says ChatGPT's memory summary "will not include everything that ChatGPT remembers."7
  2. The record won't move. Switching is improvised: Anthropic's memory-import tool has users paste a prompt into their old chatbot that asks, "List every memory you have stored about me."8
  3. Deleted isn't gone. OpenAI says fully deleting something ChatGPT knows means clearing "every source where it appears."7 From May to September 2025, a federal court ordered OpenAI to preserve chat logs "that would otherwise be deleted," including chats users asked to delete.9

Why legislation: Congress has passed privacy laws for health, financial, and children's data, but no comprehensive one; at least 19 states have written their own.10 The FTC can punish a company that breaks a privacy promise or rewrites one in secret,11 but it cannot create rights no company promised. No federal law gives Americans a right to see, move, or delete what a chatbot remembers. Companies that invite Americans' confidences owe them control of those confidences.

The Solution

A four-step staircase: each step stands alone, and each step up adds user control. Scope: consumer chatbots that keep identifiable conversations or use personal information across sessions; stronger state laws stay in force.

Step 1 — Show the whole memory. Chatbots must show users the saved facts, preferences, and inferred profiles used to personalize answers, and let them correct or delete any entry or turn memory off, noting that model training is a separate process. Today's summaries are partial; the law should require the whole record.

Step 2 — Hand over the record. On request, deliver a portable, downloadable copy of the complete chat history in human- and machine-readable formats, with dates, lawful attachments, and chosen memory entries, securely to the user or a service the user names. The bipartisan ACCESS Act would require large platforms to move user data "in a structured, commonly used, and machine-readable format."12 Users can then switch, and rivals can compete for them.

Step 3 — Make deletion real. On request, stop using and delete logs, memories, and derived profiles across personalization, advertising, profiling, and sale; pass the request to processors and recipients; and confirm completion. Legal holds and security investigations stay exempt, walled off from commercial use and explained to the user. Backups must expire on schedule and may not revive a deleted profile.

Step 4 — Ask before a confidence is sold. Require separate, revocable consent for cross-session memory, advertising, sale, and model training, plus a private mode that keeps no memory; saying no may not cost unrelated features. ChatGPT, which began testing ads in February 2026, already lets users keep past chats and memory out of ad selection;13 Meta offered no such choice. A confidence should not become a targeting signal without a yes.

Where to start: Step 1 is the floor; the largest chatbots already offer part of it. Steps 2 and 3 are the heart: take the record, then take it back.

Administration and enforcement: The FTC issues standards within 12 months, with six more months to comply. Companies answer requests within 30 days, with one disclosed extension for complex cases. The FTC and state attorneys general enforce through civil penalties and orders to finish incomplete deletions.

Risks and Mitigations

  • Total erasure: A trained model is not a database, though California already treats AI systems "capable of outputting personal information" as holding it.14 Require deletion from identifiable stores and future training sets, test for memorization, and tell users what cannot be undone rather than promise erasure. Some residue may remain.
  • Access can expose data: An export is also a target for thieves and abusive partners. Authenticate through the account, allow selective exports, protect others named in a chat, and avoid demanding government ID; no safeguard stops every account takeover.
  • Memory helps: Nothing here bans it. Users who want continuity keep it; users who want a clean slate get one.

Similar Bills

Fit measures similarity to this proposal's mechanisms: High = direct precedent; Partial = useful component with material differences; Related = adjacent approach.

Federal — 119th Congress

Proposal or bill Relevant provisions and fit Fit
H.R. 9619 — People-First Chatbot Act
Foushee (D-NC), Casar (D-TX)
Referred to committee · July 9, 2026		 §2(a)(2) requires chat logs in a "portable, readily usable, downloadable, and human and machine-readable format" and deletion on request; §2(a)(1) bars using chat logs to target ads, selling them, and profiling beyond a user's request. Direct precedent for Steps 2–4; no memory dashboard, derived-profile rule, or proof of deletion. High
S. 4199 — Youth AI Privacy Act
Markey (D-MA)
Ordered reported with a substitute · Aug. 5, 2026		 Introduced §4(b) limits personalization for known minors to data from the current session; §5 bars advertising to, profiling of, and training on minors. Strong precedent for Steps 1 and 4, but as prohibitions for minors rather than rights for all users. Compares introduced text; the substitute was not reviewed. Partial
S. 1634 — ACCESS Act of 2025
Warner (D-VA), Hawley (R-MO), Blumenthal (D-CT)
Referred to committee · May 7, 2025		 §3 requires interfaces for secure transfer of user data in a machine-readable format; §5 lets users delegate to registered agents. Transfer architecture for Step 2; covers communications platforms above 100 million monthly U.S. users, with no deletion or memory rights. Partial
H.R. 8413 — SECURE Data Act
Joyce (R-PA) + 10 Republican cosponsors
Referred to committee · April 21, 2026		 §2(a) grants rights to access, correct, delete, and port personal data and to opt out of targeted advertising and sale; §15 broadly preempts state law. Economy-wide baseline for Steps 1–3; ads and sale are opt-out rather than consent, and there are no chatbot memory rules. Partial

State

Proposal or bill Relevant provisions and fit Fit
California — CCPA, as amended
Enacted 2018; amended by Proposition 24 (2020) · In effect		 §1798.105 gives a right to delete; §1798.130(a)(2)(A) requires disclosure within 45 days in a transferable format; §1798.140(v)(4)(C) covers AI systems that can output personal information. Enacted precedent for Steps 1–3; sale and sharing are opt-out, and there is no memory dashboard. High
Maryland — SB 827 (2026)
Senate hearing scheduled March 12, 2026 · Not enacted		 Introduced §14-5104 would bar using chat logs to choose or customize ads, profiling beyond a user's prompt, and training on inputs without consent, and would cap chat-log retention at 10 years. Closest state precedent for Step 4; no export or deletion right. High
California — SB 1119 (2026)
Signed Sept. 10, 2026 · Main duties operative July 1, 2027		 §21812(d)(3) turns persistent conversational memory off by default for child users, changeable only by a parent; §21813 bars ad targeting based on a child's chats and sale of a child's data. Child-only precedent for Steps 1 and 4. Partial
Colorado — SB 21-190, Colorado Privacy Act
Signed July 7, 2021 · In effect July 1, 2023		 Rights to access, correct, delete, and obtain portable copies; opt-outs from targeted advertising, sale, and profiling; consent before processing sensitive data. General framework for Steps 1–4; no chatbot logs, memories, or inferences. Partial

What this adds: H.R. 9619 already pairs portable chat logs with deletion and an ad-targeting ban, and California now protects children's chatbot memory. This proposal reaches the memory layer itself (saved facts and inferred profiles, not only logs), requires confirmation that deletion happened, sets rules for legal holds and backups, and lets adults keep memory while refusing ads and sale. It is distinct from chatbot research access, which concerns donating logs to science, and from structural remedies, which use portability to discipline dominant firms.

Notes

  1. KFF, "KFF Tracking Poll on Health Information and Trust: Use of AI For Health Information and Advice," March 25, 2026. Self-reported; nationally representative survey of 1,343 U.S. adults, February 24–March 2, 2026. The 13% are the 41% of AI health users who uploaded personal medical information. ↩

  2. Sarah Perez, "Sam Altman warns there's no legal confidentiality when using ChatGPT as a therapist," TechCrunch, July 25, 2025. Remarks on This Past Weekend w/ Theo Von. ↩

  3. OpenAI, "Memory and new controls for ChatGPT," February 13, 2024, update of April 10, 2025: memory "now references all your past conversations." Text verified from the Internet Archive copy of September 18, 2026. ↩

  4. Google, "Gemini adds Temporary Chats and new personalization features," August 13, 2025: the past-chats setting "is on by default." ↩

  5. Meta, "Improving Your Recommendations on Our Apps With AI at Meta," October 1, 2025. Meta says it does not use conversations about religion, sexual orientation, politics, health, race or ethnicity, philosophical beliefs, or union membership to show ads. ↩

  6. Maxwell Zeff, "Meta plans to sell targeted ads based on data in your AI chats," TechCrunch, October 1, 2025: "There is no way to opt out, according to Meta." Status of the opt-out after December 2025 not re-verified. ↩

  7. OpenAI Help Center, "Memory FAQ." Text verified from the Internet Archive copy of August 20, 2026. ↩ ↩2

  8. Anthropic, "Import and export your memory from Claude," Claude Help Center, accessed September 23, 2026. ↩

  9. In re OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143, ECF 33 (S.D.N.Y. May 13, 2025) (order to "preserve and segregate all output log data that would otherwise be deleted," "whether such data might be deleted at a user's request" or under privacy law); Stipulation and Order, ECF 922 (No. 23-cv-11195, Oct. 9, 2025) (ending the ongoing obligation as of September 26, 2025; logs already preserved stay preserved, except those from the EEA, Switzerland, and the U.K.). ↩

  10. Congressional Research Service, Preemption and Privacy Law, R48667, August 29, 2025, summary page ("at least 19 states have adopted comprehensive consumer privacy laws"). ↩

  11. Staff of the FTC Office of Technology and Division of Privacy and Identity Protection, "AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive," FTC Technology Blog, February 13, 2024. Staff guidance, not a Commission rule. ↩

  12. S. 1634, ACCESS Act of 2025, 119th Cong. § 3(a) (introduced text); sponsors Warner (D-VA), Hawley (R-MO), and Blumenthal (D-CT). ↩

  13. OpenAI, "Testing ads in ChatGPT," February 9, 2026 (ads matched to "the topic of your conversation, your past chats, and past interactions with ads"); OpenAI Help Center, "Ads in ChatGPT" (toggle keeps "past chats and memory" out of ad selection). Both verified from Internet Archive copies of September 17 and 19, 2026. ↩

  14. Cal. Civ. Code § 1798.140(v)(4)(C) (personal information can exist in "artificial intelligence systems that are capable of outputting personal information"). ↩