Study: ChatGPT Saved 96% of Its User Memories Unprompted

Of 2,050 memory entries that ChatGPT saved about 80 of its users, only 84, or 4%, followed an explicit request to remember something, according to a paper accepted at the ACM Web Conference 2026 and read on arXiv on 24 October 2026. The authors are from Max Planck institutes, Ruhr University Bochum, IIT Kharagpur and Delft University of Technology. They write that “A striking 96% of memories in our dataset are created unilaterally by the conversational system, potentially shifting agency away from the user”.

A second paper, posted on arXiv on 13 September 2026 and not yet peer-reviewed, audited 179,057 conversations from 1,057 users in India, Nigeria, Brazil and Pakistan. It reports that “over 95% of profile entries are implicitly extracted without explicit user prompts or consent.”

Business Pill · CONSENT

A short Business Engineer explainer on consent: why agreeing only counts when you know what you are agreeing to. It teaches the general idea only and says nothing about the studies in this piece.

What the First Study Found

• The data. The paper, titled “The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT”, analyses 2,050 memory entries from 80 users who donated their ChatGPT data under GDPR. The memories were triggered in 1,058 conversations across 65 of the 80 participants.
• Who started the memory. The authors searched user messages for words that signal a memory request, such as remember, save and forget. They found such a request behind 84 of the 2,050 entries.
• What the memories held. The paper says memories “contain a rich mix of GDPR-defined personal data (in 28% memories) along with psychological insights about participants (in 52% memories)”. It counts 7% of memories (151) as special-category data under GDPR Article 9(1).
• How faithful they were. It also finds that “A significant majority of the memories (84%) are directly grounded in user context, indicating faithful representation of the conversations.”

Sources: arXiv 2602.01450 and arXiv 2609.14697, read 24 October 2026. The two studies use different data and definitions, so their figures are shown side by side, not combined.

What the Second Study Found

The second paper, “Vulnerabilities in Personalization: Assessing Health Privacy Risks in ChatGPT Logs and Memory”, by S M Mehedi Zaman and Md Mozammel Hoque, reports that 21.31% of the conversations it audited contain personal health data, and that 3.62% pose high-to-extreme privacy risks. The authors say they found “a stark disconnect between corporate framing and system behavior”.

They also write that “background memory synthesis selectively condenses temporary, symptom-level disclosures into permanent diagnostic traits, stripping contextual integrity and amplifying re-identification risks.” The paper is a preprint and has not been peer-reviewed.

What OpenAI Says Memory Does

OpenAI’s own post on the feature says memory first launched in April 2024 as saved memories, and that in April 2025 it introduced the first version of dreaming, “a method for ChatGPT to automatically curate memories in the background by referencing chat history.” The post adds: “Dreaming also makes it easier for memory to include context that occurs naturally in conversation, without relying on explicit requests to remember something.”

The same post announces a newer memory system built on dreaming and says “The memories synthesized by dreaming are reviewable through a summary of them made visible in the memory summary page.” It says the update “is available to Plus and Pro users in the US today, and will roll out to additional countries and Free and Go users over the coming weeks.” The first study examined saved-memory entries in donated data; it does not test the newer system.

By Our Arithmetic

By our arithmetic, 84 requested entries out of 2,050 is 4.1%, so for roughly every 24 memories ChatGPT saved in the first study, one followed a user’s request. In the second study, 3.62% of 179,057 conversations is about 6,480 conversations rated high-to-extreme privacy risk.

Connecting the Coverage

Memory has become a running thread in our coverage this week. Our Mem0 piece reported a study finding that, with Claude Opus 5.5, building Mem0’s memory store cost $25.88, against $17.84 for an entire full-context run. Our locked-chats piece reported OpenAI saying on 20 October that locked chats “let you keep selected conversations in your account while hiding them from regular chat history, pinned chats, and search.”

Our Claude finance piece reported TestingCatalog saying on 18 October that a Personal Finance feature for Claude “will work through Plaid integration”, with one screen saying “Claude reads your balances and transactions to work with your actual numbers”. Our Meta terms piece reported that the UK English edition of Meta’s Commercial Terms preview no longer contains the word “train”. Separately, Android Authority reported on 23 October that Google updated its privacy policy “to clarify that it could use information shared through Personal Intelligence to personalize ads.”

What’s new: two separate research teams, working on different users in different countries, each report that more than 95% of what ChatGPT stored about people came without a request to store it.

The Structural Read

As we read it, the two papers measure the same thing from different ends. The first counts how ChatGPT’s saved memories began, using data that 80 users donated under GDPR; the second compares 179,057 conversations with the profile entries ChatGPT built from them.

As we read it, both teams report the same order of magnitude: 96% of memories in the first study and over 95% of profile entries in the second were stored without an explicit request.

As we read it, the business question is who controls the profile. OpenAI’s post says memories synthesized by dreaming are reviewable in a memory summary page; the first paper proposes a tool it calls Attribution Shield that alerts users about potentially sensitive memory inferences.

Three Implications

FOR COMPANIES BUILDING ASSISTANTS The first paper finds 96% of memories created by the system itself and 28% holding GDPR-defined personal data, so what an assistant stores becomes a data-protection question, not only a product one.

FOR REGULATORS The first paper uses GDPR’s own definitions, Article 4(1) and Article 9(1), to classify what ChatGPT stored; it counts 7% of memories as special-category data.

FOR USERS OpenAI’s post says memories synthesized by dreaming can be reviewed in a memory summary page, and points to its Memory FAQ for controls.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework states: “The Container Moat — built by closed data loops inside a customer’s environment, nascent but the deepest position in the stack.”

As we read it, an assistant’s memory is a data loop of this kind at the level of one person: each conversation adds to a profile that only that product holds. The two papers measure how that loop fills: mostly without a request. The framework was published in July 2026 and does not discuss these studies.

What Is Not Established

The first study’s own limitations section says “Firstly, our dataset, while ecologically valid, is sourced from a small set of Prolific users primarily in the US and Europe.” Its personal-data counts rely on GPT-4o as an annotator, checked by the authors on a manual sample. The second study is a preprint. Neither paper measures whether the stored memories were used for advertising or shared outside OpenAI, and we did not find an OpenAI response to either paper in what we read.

The Bottom Line

Two studies of ChatGPT’s memory, one accepted at a major web conference and one a preprint, put the share of stored profile information saved without a detected user request at 96% and over 95%. OpenAI’s post says saved memories “relied on strong cues to decide when to trigger memory”, while the first study, which examined saved-memory entries, found an explicit request behind 84 of 2,050.

OpenAI’s post also says its dreaming system, which the first study did not test, “makes it easier for memory to include context that occurs naturally in conversation, without relying on explicit requests to remember something.”