We remember only a tiny fraction of our experiences. While we often say "I forgot X", most memories are not really forgotten—with the right cue, in the right setting, dormant memories can return to us spontaneously, effortlessly. However, we cannot recall most of our memories at will. For example, try recalling the names of all your classmates in first grade. This type of remembering, known as free recall, is quite challenging. Now, suppose instead that you only need to pick out your classmates' names from a list that includes your classmates among other names. You are likely to recognize many more of your classmates' names from this list than you can remember off the cuff. This is known as cued recall, and it is much easier than free recall.
Figure 1: Most personal memories are hard-to-recall "dark matter" of memory. All the information there is to know may be classified into three kinds. First is world knowledge, such as the name of the longest river in the world. Second is personal memories. Personal memories include some world knowledge, such as who the Secretary-General of the UN is. Nevertheless, many of our personal memories—say, the gift we bought for our parents with our first paycheck—are not world knowledge. Third is what we can easily remember from our personal memories. The vast majority of personal memories are hard to recall; they form the dark matter of memory.
Memories that are not really lost, but that are only accessible via the right cues, are the dark matter of memory. Like dark matter in the universe, the dark matter of memory is massive, ubiquitous, yet not directly observable (Figure 1). To begin to characterize the dark matter of memory, we collected a dataset of everyday memory questions in the wild (QITW)—questions about information people once knew, but can no longer recall. These questions are the most noticeable penumbra of the dark matter of memory.
We asked 134 participants (46% female, ages 18–80, mean age: 41) to note down memory-related questions they had during daily activities. Memory-related questions were described as occasions when "you need to search for something you once knew." Detailed task instructions are included in Appendix 1. All participants were fluent in English; 86% were native speakers. The participants' occupations are summarized in Table 1. We recruited participants via the Prolific crowd-sourcing data collection platform. Participants indicated their written consent for completing the study and were compensated for their efforts.
From a total of 3,609 questions, we excluded questions that could be answered with just world knowledge, obtaining 1,940 personal memory questions. Participants asked a wide variety of questions. Examples include "Who is …?", "When did I last meet with …?", "What is my username and password to log in to …?", "Where is the folder with the current slide deck from …?", and "What did my manager ask me to do about …?". Additional example questions are shown in Appendix 2.
Mean 41, Median 41, Range 18-80, Mostly 25-54.
Word frequency highlights key themes including social interactions and time
We visualized the frequency of words across questions (Figure 2), excluding articles (e.g., "the"), verbs (e.g., "did"), and adjectives (e.g., "many"). Word frequency reveals many of the key themes in the data. People frequently asked about names, people ("name," "friend"), and time ("time," "year," "week," "date," "today"). People also often asked about exact details such as passwords, phone numbers, emails, birthdays, folders, and documents.
Figure 2: A word cloud of memory questions reveals themes such as names, people, time, and exact information. The size of each word is proportional to its frequency of occurrence among the questions.
"What" questions were the most common
We classified the questions based on their primary intent: "What," "When," "Where," "How," "Who," "Why," or "Whether." This classification was based on not just syntax but also meaning using an LLM (gpt-5.2; the prompt is shown in Appendix 3). For example, "I wonder what is the name of the TV series [...]" was classified as a "What" question, while "What was the name of the person" was classified as a "Who" question. "What" questions were the most common by far, comprising approximately 40% of all questions (Figure 3).
Figure 3. Participants most frequently asked "What" questions. We classified questions by intent into "What," "When," "Whether," "Where," "How," "Who," or "Why." Each small dot indicates the fraction of questions a participant asked in a category. The size of each dot is proportional to the number of questions. Diamonds with whiskers indicate the across-subject averages with 95th-percentile confidence intervals.
Participants most frequently asked about their own past actions
We further categorized the QITWs into 18 detailed categories (including an "unassigned" null category for ambiguous questions). The categories were created by manually reviewing a representative subset of questions, and then applying the categorization scheme to the remaining questions using an LLM (gpt-5.2; the prompt is shown in Appendix 3). The most common questions were about participants' own past actions. Of these, most questions were about the recent past—hours or days before posing the question. The next most common questions were about contact information, followed by questions about schedules, object locations, and tasks.
Figure 4. People typically asked questions in a consistent mix of categories, most commonly about their own past actions. Questions about recent past actions (e.g., "What was the name of the cafe I visited last weekend?") were more common than questions about actions further in the past (e.g., "What present did I give my daughter on her 10th birthday?"). The next most frequent questions included contacts, plans, misplaced items, followed by personally relevant world-knowledge questions (e.g., tip-of-tongue questions). This figure follows the format of Figure 3.
Participants most often asked memory questions when working and planning events
The participants also recorded what they were doing when they had each memory question. These activities were grouped into 16 activity categories (Figure 5). About a fifth of the questions arose while people were working, followed by planning, socializing, and web browsing. The frequency of activities reflects both when memory questions occur and the general frequency of activities. Nevertheless, the range of activities demonstrate that memory needs pervade all aspects of life.
The type of memory question participants asked correlated strongly with what they were doing (Figure 6). We quantified the correlation between activity type and question type using odds ratios. For example, participants were 11 times more likely to ask about locations when traveling and 3 times more likely to ask about schedules when making plans. In these examples, the activity type congrues with the question category. Such congruent cases are plotted along the pseudo-diagonal in Figure 6 and account for a large portion of the questions. There were also notable cases outside of the congruent pairs. For example, watching TV and reading strongly triggered tip-of-tongue questions (6.2× and 5.8× more likely), while browsing the internet broadly triggered questions about account IDs (6.4×), passwords (4.3×), and trivia (2.4×).
Figure 5. Participants needed to recall information from memory across multiple activities, particularly when working and planning events. Each small dot indicates the fraction of questions a participant asked during an activity. The size of each dot indicates the number of questions that participants asked during that activity. Diamonds with whiskers indicate the across-subject averages with 95th-percentile confidence intervals. The most common activities when people asked memory questions were working (19.8%), planning/organizing (12.8%), and socializing (11.0%).
Figure 6. What participants were doing correlated strongly with the type of information they sought. Each row is one of 14 activity categories from Figure 5, and each column is one of the memory question categories from Figure 4. The rows and columns are sorted to align congruent pairs (where the activity naturally matches the question category; highlighted in orange borders) along the diagonal. Each cell shows the odds ratio, i.e., the odds that a particular type of memory question arises during a particular activity, relative to the odds that the question type comes up during other activities. The color scale is logarithmic and centered at 1× (white); blue indicates enrichment (higher odds), and grey indicates depletion (lower odds). Congruent pairs are highly enriched. For example, when cooking, people are 13.7× more likely to ask about recipes; people traveling are 11× more likely to ask about locations; and people planning are 3× more likely to ask about schedules. Off the diagonal, watching TV and reading both strongly trigger tip-of-tongue questions (6.2× and 5.8×), while browsing the internet leads to broad enrichment across questions about account IDs (6.4×), passwords (4.3×), and trivia (2.4×).
Memories give color to our unique identities, yet most of our memories are usually hidden and inaccessible (Figure 1). To begin shedding light on the dark matter of memory, we collected questions that people ask daily about their memories, at work and in life. We refer to these questions as memory questions in the wild.
A large body of laboratory-based experiments has asked participants to memorize and recall information [2, 6, 8]. For example, studies have presented participants with lists of words and then asked them to recall those words after certain time intervals. By design, these experiments are impersonal, unlike the organic, deeply rooted memory questions people ask themselves daily. More recent studies have characterized memories of real-life experiences such as movie watching [4], city walks, and museum visits [5]. While these naturalistic experiments better capture the encoding and recall of real-world information, they remain specific to controlled experimental conditions that do not always reflect our actual memory needs. In contrast, here we characterized people's day-to-day memory needs.
These memory questions in the wild confirm several intuitions and reveal unexpected insights. People often asked about contact information such as emails and phone numbers, schedules, tasks, and passwords—categories for which we have invented dedicated tools, from physical planners and calendars to digital notes and keychains. Meanwhile, the questions reveal prevalent memory needs that remain ill-served. The most common questions ask about participants' own actions, among which the recent past ("What was the name of the cafe I visited last weekend?") features more often than the deeper past ("What present did I give my daughter on her 10th birthday?"). Participants tended to ask about the recent past perhaps because it is more relevant to them at the present moment. However, this recency bias can also arise whenever older events are harder to ask about—because those older events fade into the dark matter of memory. For example, a participant may not ask about the cafe visited on a certain weekend last year because they may not remember having visited a cafe on that weekend. Old memories are nevertheless precious, as diaries and autobiographies attest.
Memory questions pervade a broad range of activities throughout the day. When working, people ask about professional task knowledge. When socializing, people ask about acquaintances and contacts. Even when watching TV and movies, people ask tip-of-tongue questions about shows, films, and more. Besides the ongoing activity, myriad other factors, such as age, health, profession, company, and the time of day, influence our memories. Thus, this study is only an initial step toward characterizing memory questions in the wild. Future work may focus on work versus personal life, different professions, age groups, and more, leading to an ever more thorough and systematic understanding of people's memory needs.
An important direction of future work is to characterize the needs of people with progressive degrees of memory loss. Understanding those most in need of memory assistance can help improve their quality of life, and can be a crucial first step to develop a technology that improves memory for all.
The ubiquity of memory questions—what I know I knew (but can't remember)—only begins to peer into the dark matter of memory—what I don't know I knew. We can only realize, ask about, and search for information we partially remember. It is impossible to ask about events and topics we remember nothing about. By understanding people's observable memory needs and building tools to meet them, we are making inroads into revealing ever more of the hidden dark matter of memory.
What is QITW?
Questions in the wild (QITW) are all your memory-related questions:
There are no wrong questions; there are no small questions. Every question that should be in your memory but is not belongs in QITW.
Instructions
We are conducting a research project about memory. We are collecting data about what memories you need but can't remember throughout the day.
We would like to know about every time you need to search for something you once knew. Please be specific when describing your questions.
Please record these questions here as you come across them throughout the day.
This appendix shows example questions submitted by participants, categorized in two ways: W-questions (Figure 3) and 18 detailed question categories (Figure 4). The questions are copied verbatim, except when modified to omit personal information.
W-questions
What/which
When
How
If/whether
Where
Who
Why
Eighteen (18) detailed categories
Recent past actions
Autobiographical past
Schedule/time
Contacts (general information)
Tasks/to-do's
Object location
Tip-of-tongue
Trivial/factual
Professional task knowledge
Digital file location/URLs
Procedures/how-to/recipes
Passwords
Personal task knowledge
Household objects and state
Others' contact/ID
Own accounts/IDs
Event physical location
Unassigned
W-questions
You classify personal memory questions by their W-type question intent.
This scheme is meant for messy, real-world language: implicit questions, missing entities, vague references, and partial context are all common. Use both the question text (including explanatory comments inside it) and the contextDoing field. If a question does not fit a single category with high confidence, use unassigned.
Categories (use labels exactly)
Rules
Eighteen detailed categories
You classify personal memory questions into specific categories.
Input fields
Task
Choose exactly one category label for each question.
Categories (use labels exactly)
Core rules
Tie-break guidance (still prefer unassigned if uncertain)