Two studies, two sides of the same feature
Companion apps sell continuity: a character that recalls your week, your worries and your name. Researchers have mostly asked whether that makes people happier or lonelier. These two papers ask something narrower and more practical: does remembering make people share more, and what happens to what they share?
| “Tailored to you” | “Vulnerabilities in Personalization” | |
|---|---|---|
| Posted | 17 September 2026 (arXiv:2609.20077) | 13 September 2026 (arXiv:2609.14697) |
| Team | Google DeepMind (nine authors) | Rutgers University (two authors) |
| Design | Preregistered randomised trial, five consecutive days | Audit of donated real-world chat histories and memory entries |
| Scale | 992 participants, 4,960 conversations | 1,057 users, 179,057 conversations in India, Pakistan, Brazil and Nigeria |
| Main question | How personalisation changes trust, disclosure and regret | What a live memory system extracts and keeps |
The DeepMind trial: memory feels natural, and people open up
Participants, recruited through Prolific, spent 15 to 20 minutes a day talking with a model about everyday relationship topics such as setting boundaries or repairing a strained friendship. They were randomly assigned to one of three versions: a control model with no personal information; a memory model that received a running summary of their earlier sessions; or a survey model that received a persona summary written from an intake questionnaire.
Most of the change came from repetition, not personalisation
Over the five days, people in every group felt closer to the model and rated it as more useful and more competent. Personalisation did not speed that up. In the authors’ words, several shifts were driven primarily by repeated exposure rather than personalisation itself — a reminder that simply talking to a chatbot every day changes how people relate to it.
Where memory did make a difference
- Less creepy. The memory version was rated significantly less creepy than the control. The survey version was not.
- More detail shared. Self-reported depth of sharing did not differ between groups, but the number of personal details people actually disclosed rose faster across sessions with the memory model. What people felt they shared and what they shared did not line up.
- More regret with surveys. People whose model was primed with their questionnaire answers reported marginally higher regret about what they had shared. Memory did not raise regret.
The Rutgers audit: what memory actually keeps
The second study looks at the other end of the pipeline. Users in four countries donated their full ChatGPT export files through a consented, ethics-approved process, and the researchers analysed both the conversations and the memory entries ChatGPT had built from them.
- 21.31% of conversations contained personal health information, and 3.62% carried high-to-extreme privacy risk, such as stigmatised conditions, direct identifiers or precise locations.
- Over 95% of memory entries were extracted implicitly, without an explicit user prompt or real-time confirmation — despite documentation that framed memory as user-directed.
- Memory tended to condense temporary, symptom-level remarks into permanent diagnostic traits. A one-off question about a symptom could become a standing line in a profile.
The authors note that this is a ChatGPT finding: they could not run the same audit on other assistants, and they argue the risk follows from specific design choices rather than from memory as such.
Why the two findings belong together
Read side by side, the studies describe one loop. Memory makes a model feel familiar and less intrusive, people share more with it, and an automatic memory system then turns what they share into lasting records — often without asking. Neither paper tested a companion app, but companion apps lean on memory more heavily than general assistants, and the conversations are often more personal. Our explainers on persona memory and why companions forget cover how these systems are usually built.
A memory checklist for any companion app
| Check | Why the research says it matters | What to look for |
|---|---|---|
| Can you see what it remembers? | Most entries in the audit were created without the user asking | A readable list of saved memories, not just a toggle |
| Does it confirm before saving? | Real-time confirmation was largely absent | A notice or prompt when a new memory is written |
| Can you delete one memory? | Passing remarks can harden into permanent traits | Per-item delete and edit, not only “wipe everything” |
| Can you exclude a topic? | Health was the most common sensitive category | Topic-level or temporary chat modes that skip memory |
| Are you sharing more than you think? | Disclosures rose with memory while self-reports stayed flat | Review the memory list after a few weeks and compare |
For the wider settings audit, including data export and deletion, see our AI companion privacy checklist.
What these studies cannot tell you
- Both are preprints and have not completed peer review.
- The DeepMind trial lasted five days, used assigned topics and personalised only through the prompt; longer, free-form use may differ.
- The Rutgers audit covers ChatGPT only, used an LLM-assisted coding pipeline checked by human coders, and its donors were people able to export their data, which the authors flag as a selection bias.
- Neither study measured harm from a leak or misuse; they measure what is shared and what is kept.
The bottom line
Memory is not neutral plumbing. One study shows it makes people more comfortable and more forthcoming; the other shows how much of that openness can end up stored without a clear moment of consent. When you compare companion apps, the useful question is not whether they remember, but whether you can see, confirm and remove what they remember.
Related on CompanionRank
AI Companion App Privacy Settings & Safety Checklist 2026 What Is Persona Memory in AI Companion Apps? Memory and Continuity: Why AI Companions Forget Things How to Export Your AI Companion Chat HistoryFrequently asked questions
Does AI memory make people share more personal information?
In a preregistered Google DeepMind trial with 992 participants over five days, people talking to a model that remembered earlier sessions disclosed personal details at a faster-growing rate than those using a model without memory, even though their self-reported level of sharing did not differ between groups.
Do people find AI that remembers them creepy?
Not in this trial. The memory-based model was rated significantly less creepy than the non-personalised control. A version primed with an intake survey was not rated differently, but its users reported marginally higher regret about what they had shared.
How much of ChatGPT memory is saved without asking?
A Rutgers University audit of 179,057 donated ChatGPT conversations from 1,057 users found that over 95% of memory entries were extracted implicitly, without an explicit user prompt or real-time confirmation.
How often do people share health information with chatbots?
In the same Rutgers audit, 21.31% of conversations contained personal health data and 3.62% carried high-to-extreme privacy risk, such as stigmatised conditions, direct identifiers or precise locations.
Did these studies test companion apps like Replika or Character.AI?
No. The DeepMind trial used language models personalised through the prompt, and the Rutgers audit covered ChatGPT only. Both are preprints that have not completed peer review, so their results are best read as signals for how memory features behave in general.