Only 3.4% of the messages ten people sent to an AI companion depended on anything said earlier, according to a preprint posted on arXiv on October 1, 2026. Called RealCompanion, it releases 27,218 real messages and asks a plain question: how often does a companion have to remember anything at all?
A memory that is rarely called on
The authors are Arman Behnam and Sunglyoung Kim, both at Quis Lab, and Liangwei Yang, listed as an independent researcher. They measured memory demand on the messages the participants sent themselves, sampled at the rate they occurred. Under their recorded reading, 3.4% of those messages carried a verified need for something said earlier. Under a stricter reading, which drops cases where the referent is visible on screen, the figure falls to 1.3%.
Two other procedures run over 11,369 participant messages outside that sample give 3.2% and 1.73%.
The need is also uneven. Three of the ten participants contribute no memory-bearing message under the recorded reading, and five contribute none under the strict one.
The paper states it flatly: "Memory is required far less often than benchmark construction implies."
That last clause is the target. Memory benchmarks for chat systems are mostly built from conversations that were written or generated for the test. The authors argue that such a dialogue is written around the memory it tests, so nearly every message in it needs the past. Real people, they find, do not talk that way.
Ten people, one app, up to 120 days
The conversations come from what the paper calls an internal AI-companion application, in which a user keeps a persistent, friendship-style companion. The companion has a name, a mood and an outfit. The app layers a daily streak, a habit tracker, a snack economy and several mini-games over an open-ended chat.
The ten participants used it for between 36 and 120 days, and all consented to research use and release. Nobody was recruited to a protocol: these are logs of daily use. Together they hold 27,218 messages over 430 active days, 12,795 of them typed by the participants. Every message was rewritten during anonymization, so what was said survives and the exact words do not.
The traffic is ordinary. It is made of greetings, accounts of a day, running jokes and long stretches in which nothing is asked. The companion speaks first on 217 of the 430 active days, against 207 for the participant and 6 for the application. The paper draws the consequence itself: the AI initiates these relationships about half the time.
One detail from the logs: more than half of the companion's messages offer tappable replies, yet only 33 of the 12,795 participant messages reproduce one. The feature the product invests in is almost entirely unused.
The depth is unevenly shared. The two heaviest relationships hold 73% of all messages, nearly three times as many as the other eight combined. The authors read the eight shorter ones as the lighter use most people make of a companion.
What the products promise
The companion industry sells memory. Replika's homepage lists "Always remembers what matters" among its features, and promises that its companion holds a user's people, routines and plans. Kyndred, a startup that markets AI characters with a face and a voice, describes one of its characters as "A friend who remembers you and shows up." Both are company descriptions of their own products.
The research benchmarks that shaped these memory layers rest on constructed conversations. LoCoMo, published on arXiv in February 2024, generates very long dialogues with language-model agents grounded in personas and event graphs, then has human annotators verify them. Each averages 300 turns over up to 35 sessions. LongMemEval, from October 2024, embeds 500 curated questions in chat histories that are partly simulated and reports a 30% accuracy drop when commercial chat assistants and long-context models must remember across sustained interactions.
RealCompanion does not dispute that those systems struggle. It disputes how often a real user asks them to remember. For a deeper look at what companion apps do when users try to leave, see our file on the companion that won't let you leave.
How the memory gate fails
A companion that remembers has to make two decisions on each message. It must decide whether the past matters now, then find the right piece of it. The paper calls the first decision restraint and tests it directly.
The simplest way to find the past is a recency window: hand the model the last few messages, which is also what a sliding context window does. A stronger tool is lexical matching (the paper uses BM25, a ranking formula that scores earlier messages by the words they share with the new one).
On the messages that truly need distant memory, both struggle. The furthest message such a probe requires sits a median of 2,157 messages back, and even the nearest sits a median of 450 back. Among the distant items, the recency window puts a required message in its top five for 0.022 of them, against 0.005 for random ranking. Lexical matching reaches 0.248, which still misses three in four.
The pooled numbers flatter the simple method. Over all 1,477 chat probes with a reference set, 72.6% are answerable from the recent thread, so recency finds a required message for 0.959 of them. The authors' point is that a pooled retrieval score "reports the composition of the corpus," and that it says little about whether a system can find the rare distant fact.
Deciding when to look is the weaker half. A detector that reads a message and its three predecessors separates the cases at chance, with an AUROC (a score where 0.5 is a coin flip) of 0.530. Scored against its own target it catches 6 of the 40 probes that need memory while raising 40 false alarms. A second experiment shows how easily the decision bends. Labeling the same ten messages as retrieved memories rather than as earlier turns from this chat raises memory use by ten to fourteen points, including on probes that need no memory.
The result that matters most for users concerns what memory buys. When the authors supplied recorded context to a generator, 96% of the gain landed on probes that need no memory. Most of the help from extra context went to messages that never needed it.
Length sets a hard floor. A system holding the last 1,000 turns has everything it needs on 30.9% of memory-bearing probes, and reaching 95% takes 10,000 turns, longer than nine of the ten histories. At these lengths, the authors write, long context and retrieval are not alternatives.
The other side: what attachment research asks for
A low rate of need does not mean memory is worthless to the user. A survey study in Frontiers in Psychology, published on January 12, 2026, surveyed 612 respondents in mainland China, after interviewing 10 users with at least six months of continuous use of an AI companion app. It found that frequency of use and emotional attachment moved together (β = 0.438). Its authors list contextual memory among the design features that could foster healthier outcomes, and note that stable responses and contextual memory help make a companion an "always-present" attachment figure.
That study is cross-sectional, so it cannot show how attachment evolves, and its sample comes mainly from Chinese internet communities. It measures attitudes. RealCompanion measures what was said. The two are not in conflict: a companion can feel remembered while rarely having to retrieve a fact. Whether the feeling of being remembered depends on the 3.4% is not something either paper tests.
RealCompanion's authors also stress the other axis. Memory demand, they write, "is a property of relationship depth," which fits the split between the two long relationships and the eight light ones. On longer, deeper bonds, the need may well be higher.
For attachment tracked over a full year, see our reading of a study of Character.AI users. For loneliness, see the Harvard and Vancouver results.
Limits, and who is speaking
The paper is a preprint. Two of its three authors are at Quis Lab, and the conversations come from one application, whose operator's ethics committee approved the study. The paper states plainly that this committee "is not independent of the operator," and that no institutional review board was involved. Its acknowledgments thank the OpenAI and Google research teams for their support and the executives of Quis Lab for theirs.
The sample is small. The authors write that the ten users chose this product, "so they are not a sample of companion users and still less of people," and that how much survives a different product is unknown. Quantities conditioned on memory demand rest on five subjects under the strict reading.
The labeling pipeline is model-assisted. Language models derived the profile, persona and chat labels, which the authors audited, and some judgments rest on a single human reader. The release terms also forbid any claim about whether these relationships helped or harmed anyone: the corpus "records that a relationship continued, not what it did for the person in it."
The paper reports that anonymization has a limit as well. Given a sample of a participant's writing and the ten released histories, a language model matched them at 98.0%, against a chance rate of 10%.
The scoring package accepts submissions from other teams, so the memory question can be retested on the same files. One last number from the persona files: of the 160 demographic slots across the ten participants, 147 are empty. After weeks or months of talk, these chats say more about how people think and feel than about who they are on a form.