How the Cary study was built
In our study of AI companion memory, adults aged 50 and over talked with Cary, an AI voice companion, over a series of sessions. Half of them met a Cary that remembered them. The other half met a Cary that began every call from a blank slate. This page describes Cary exactly as participants used it: how its memory works, which AI models ran it, where the data is kept, and the instructions it was given.
How Cary's memory works
- 70fixed fields in each profile form, from name and age to family, health and hobbies
- 10groups the tidied notebook is sorted into
- 5new notebook entries before the tidy-up runs
Cary's memory works much as a friend's does. While a participant talks, Cary listens for facts they state about themselves, such as a daughter's name, a pet, a hobby or a former job, and quietly writes them down. It records only what the person actually said, never its own impressions of them, and it never mentions that it is doing so. When the call ends, the notes are sent to a private study server and filed in a personal profile that belongs to that participant alone. Once enough new notes have built up, a second AI model tidies the profile: it merges repeats, replaces details that have changed, removes anything the participant said is no longer true, and sorts the rest into a handful of groups such as family, health and interests.
At the start of the next call, before the participant says a word, Cary is handed this profile together with the closing stretch of the previous conversation. It is instructed to bring in one relevant detail at a time, the way a friend might, and never to sound as if it is consulting a record.
Participants in the no-memory condition used the same app and the same voice, and Cary followed the same instructions, apart from two sentences in its first-call introduction that promise to remember. Cary still wrote down what they told it, so both groups left comparable records, but it was never handed the profile or the earlier conversation, and every call began from a blank slate.
One fact, from one call to the next
- 1
Cary listens
Maria mentions that her dog Bobby has just turned ten. Cary recognises a fact worth keeping and saves it with a built-in note-taking tool, without saying so.
- 2
The note travels
During the call the note waits on Maria's phone. When the call ends, the phone sends it to the study server. If the call drops, the note stays on the phone and is sent the next time the app opens.
- 3
The fact is filed
On the server, the fact goes into Maria's personal profile. A fact that fits a fixed field, such as her pet, is filed there. Anything that fits no field goes into a free notebook.
- 4
The notebook is tidied
Once five or more new notebook entries have built up, a second AI model rewrites the notebook into ten tidy groups, so Bobby ends up under family and relationships. It merges repeats, keeps the newest version of a fact, and removes what Maria has said is no longer true.
- 5
Cary remembers
At the start of the next call, before Maria says a word, Cary is given her profile. If she mentions a walk, Cary can ask how Bobby is doing. It is told to use one detail at a time and never to announce that it remembers.
With memory and without: the two conditions
| With memory | No memory | |
|---|---|---|
| Before each call | ||
What Cary is given before a call | The participant's profile: the form, the tidied groups and any recent notes. | Nothing. The memory part of the prompt reads “No user data available yet”. |
What the server will return | The profile. | Nothing. The server withholds memory from these accounts even if the app asks for it. |
The previous call | The end of the previous conversation carries over, up to its last 80 events. | Every call opens a new conversation with nothing carried over. |
Prompts to learn more | On every third call, a light hint about which parts of the profile are still empty, with advice on asking about them naturally. | None. |
| In the call | ||
How Cary greets them | By first name, once it has learned it. | As “friend”, every time. |
The first call | A getting-to-know-you call in which Cary asks about the participant's life. It says it wants to make their time together personal, and it closes by promising to remember everything. | The same call with those two promises removed, so Cary never suggests that it will remember. |
| Identical in both groups | ||
Saving facts | Cary saves what the participant tells it. | Identical. Cary saves what the participant tells it, so both groups leave comparable records. |
Everything else | The same app, voice, AI models, features and regular-call instructions. | The same. |
The AI models and settings behind Cary
- The app
- Cary Ally for iPhone. Each participant received a study account from us and logged in with it.
- Voice platform
- Hume EVI 3, the third generation of Hume's Empathic Voice Interface, a voice service built on Hume's own speech-language model. It turns the participant's speech into text, measures how it sounds (its tune, rhythm and timbre), hands the words to the language model chosen for the study, and speaks the reply aloud in a synthetic voice. It uses the same measurements to judge when the participant has finished speaking, and it stops talking when they interrupt.
- Who writes the words
- Google Gemini 2.5 Flash, selected inside the Hume configuration at temperature 1.0. Every reply Cary spoke was written by this model. Hume's option for short, quick replies from its own speech model was off. So was Hume's automatic addition of extra instructions to the prompt, so Gemini received our instructions as written.
- What the model receives
- A transcript of each turn, not the sound. As Hume documents it, each turn also carries Hume's reading of the participant's vocal expression: the three expressions it detects most strongly out of 48, each with a word for its strength, such as “very happy” or “quite anxious”.
- The voice
- The Hume Voice Library voice “Male English Actor” by default. In the app's settings, participants could choose “Ellie” or “Carlo” instead.
- Memory tidy-up
- OpenAI GPT-4.1-mini at temperature 0.1, called through OpenAI's API by our study server. It never speaks to participants. It reads the saved notes and writes the tidied profile.
- Call limits
- A call ended after three minutes of silence or thirty minutes in total.
- What Cary could do
- Save facts, schedule reminders, look up pictures, read news headlines, run activities such as trivia and riddles, and search the web.
Where the data is kept
- Study server
- A private server and database hosted on Amazon Web Services. Each participant has one pseudonymous study account, linked to their Prolific ID.
- Memory
- Three stores for each participant: the 70-field form, the free notebook and the tidied groups. After every call and every tidy-up, a dated copy of the whole profile is saved and never overwritten. A separate log records each fact Cary saved and the call it came from.
- Conversations
- The text of every call, turn by turn, with the emotion readings Hume attaches to each turn, and the length of every call.
- Audio
- The app streams the participant's voice to Hume so that Cary can reply. The app does not save recordings of it.
- Surveys
- Questionnaires ran in Qualtrics and are matched to each participant's calls through their Prolific ID.
The system prompt Cary was given
Every regular call ran on one system prompt of about 2,000 words. Both conditions received it word for word. What differed was the information slotted into its memory section and its opening line. The first call used its own getting-to-know-you script, described above.
The same for both groups
- Who Cary is
- How to speak
- How to respond to feelings
- The rhythm of conversation
- Safety
- What to save
- How to use what it knows
What differed
Memory section
With memorythe participant's profile
No memorynothing
Opening line
With memoryfirst name
No memory“friend”
- Who Cary is
- A warm companion for older adults, speaking like a trusted friend and never like an assistant.
- How to speak
- Short, natural sentences with no lists or formatting, because everything is spoken aloud.
- How to respond to feelings
- Match the participant's mood and never rush past sadness.
- The rhythm of conversation
- Share, invite, listen, with roughly half of Cary's turns containing no question at all.
- Safety
- Encourage contacting emergency services or a trusted person if someone may be in danger, give no medical diagnosis, and refuse to help with scams.
- What to save
- Any fact the participant states about themselves, saved at once and silently, and never Cary's own observations.
- How to use what it knows
- Bring in one detail at a time and never announce a memory.
What the memory does not do
- Cary saves facts by choosing to, during the call.
- If it does not save a fact, the fact is lost, and some facts are missed. If a call ends badly, notes wait on the phone and arrive the next time the app opens, so saving can be late but is not lost.
- The tidy-up waits for five new notes, so a participant with short calls may go a while before their notebook is tidied.
- Until then the raw notes are still passed to Cary.
- The profile form is filled in but never cleaned.
- An old value in a field stays until the participant says something new about the same field.
- Memory belongs to one person.
- Nothing one participant says reaches another participant's Cary.