The Cue-Recalled Autobiographical Memory test is a computerized interactive test presented in web-browser format. It collects counts of the number of details (elements) within categories (features) that compose naturalistically word-cued AMs dated to specific life periods. Complete specifications of the test and instruction provided to participants have been previously reported² (for full detail see Gardner et al., 2012). Here, each section of the test is briefly described.

Prior to eliciting AMs, CRAM collects demographic information for each participant. Subjects are then presented the following definition of AM and subsequent instruction:

“Autobiographical memories are recollections of past episodes directly experienced by the subject. These memories should be of a brief, self-consistent episode of your life. An episode can be as short as a single snapshot and up to a few seconds long. … If the memory you think of refers to a typical and repeated episode that happened regularly or multiple times in your life, you can use it only if you can fixate on a specific individual event. If you can only recall the generic (repeated) event, look for another memory.”

The test is designed to probe memory of unique personally experienced episodes, for example, memory for a conversation one had with friends at a recent dinner party, or memory of meeting one’s spouse. This type of memory is considered episodic and has been frequently contrasted with memory for factual information about the world or oneself (semantic: Tulving, 1972, 1985).

Naturalistic word-cues are then presented to elicit memories. The participant reads through a list of seven words and labels the first recollection retrieved, for subsequent identification. Subjects are further instructed that the cued AM does not necessarily have to relate to any one word or to the entire list of words, but rather is “the first autobiographical memory that the words bring to mind.” The list of word-cues is randomly selected from the British National Corpus³, a compilation of 100 million written and spoken words (see Gardner et al., 2012 for word processing details). Thus, this procedure provides cues that are presented proportionally to word frequencies observed in everyday settings. For example, the word “waiting” appears in the corpus 474 times whereas the word “yard” appears 71 times. Therefore, “waiting” is ∼6.7 times more likely than “yard” to be presented as a word cue. As the words are sampled randomly from the corpus each time a list is generated, the specific words selected may be dramatically different for any two presentations within and across subjects.

Once AMs are cued and labeled, participants are presented with their AM labels one by one to date each memory. Specifically, the participant places each memory into one of 10 temporal bins, which segment his or her life span into 10 equal intervals (a similar binning procedure has been used previously: see McCormack, 1979; Howes and Katz, 1992; Gardner et al., 2012). If necessary, up to three temporal bins could be assigned to a single AM. To increase dating accuracy, the temporal range associated with each bin (i.e., the cutoff values computed from the subject’s age) is presented in terms of time from the present, age of the participant, and month and year. However, as the size of a given temporal bin increases linearly with the participant’s age at the time of testing (e.g., the size of a bin for a 20 year old is half that for a 40 year old), this procedure also reduces the temporal resolution associated with a memory in older (relative to younger) subjects.

The age of the participant at the time of a recalled event is estimated as the mid-point of its assigned temporal bin. The retrieval probability of an AM falling within each bin or age range is computed by dividing the number of AMs dated to each temporal period by the total number of dated AMs. Collectively, these measures are used to construct the temporal (life span) distribution of AMs. In this work, Recent AMs are defined as memories of events that occurred within the most recent 10 years of life; Remote AMs are defined as memories of events that occurred more than 10 years from the present moment. Overall, ∼4% of AMs were assigned to more than one temporal bin (presumably resulting from a lack of confidence that the AM occurred within the temporal range associated with a single bin). This proportion mildly increased for Remote AMs (Remote: 5.0%; Recent: 3.7%, p < 0.001), suggesting a reduction in a given subject’s confidence to date these older episodes. In addition, younger subjects dated AMs to multiple bins slightly more frequently (4.0%) than older subjects (3.0%; p < 0.05), potentially due to the age-dependent nature of a given bin’s temporal interval size (the temporal range associated with a bin increases proportionally with the subject’s age at the time of testing).

After AMs are dated, participants are once again presented with their AM labels to score the content associated with a selected memory. In particular, participants are instructed to count the number of details remembered within each of eight categories, i.e., Things (objects), Feelings (emotional details), People (unique individuals), Places (spatial details), Times (temporal details), Episodes (temporally linked events), Contexts (other contextual details), and Details (all remaining details, including actions). Participants provide counts of details rather than event descriptions. The order in which each category is presented is randomized for each person but fixed across all AMs for a given individual. The exact definitions of these eight categories were previously reported (Gardner et al., 2012) and are available at http://cramtest.info. In addition, CRAM provides, through clickable links, additional examples of what constitutes a detail within a given category and general scoring guidance. Every detail category is called a “feature,” each reported detail associated with a given feature is referred to as an “element,” and the summed number of elements across all features for a given memory is called “total content.”

The Cue-Recalled Autobiographical Memory test was completed locally under experimenter supervision or remotely over the Internet⁴. From each subject who completed testing in person, 30 AMs were cued and dated, of which a subset of 10 was scored for content (see Gardner et al., 2012; the first two AMs were considered practice and not analyzed). AMs were selected for scoring to maximize life span coverage in the entire dataset. Specifically, a single AM was scored from each life span bin represented by a participant. From the participant’s remaining AMs, memories were selected in order from the least to most represented bin (in the entire dataset across all participants) until 10 memories were scored in total; only for an initial sample of subjects (n = 110) the number of scored AMs was instead a function of the number of temporal bins represented.

In contrast to in-person testing, CRAM’s online protocol offers subjects the choice of several test options (i.e., Atomic, Mini, Extended, and Full tests) which differ according to the number of AMs cued, dated, and scored. These options are included to promote test completion by suiting a wide range of subjects who may vary in their commitment and eagerness to participate. The Atomic test cues one AM which is dated and scored for content. At the end of the Atomic test, subjects are asked if they would like to complete the Mini or Full test. The Mini test cues five AMs, which are each dated and scored. At the end of the Mini test, subjects are invited to extend the test. If a subject agrees, an additional fifteen AMs are cued and dated, five of which are scored; this option is categorized as the Extended test (which in total cues and dates 20 AMs, and scores 10 for content). The Full test cues 20 AMs, each of which is dated (only nineteen subjects completed instead a different Full test version which cued and dated 30 AMs as outlined in the in-person protocol). Subsequently, a subset of 10 memories is scored for content. Memory selection for scoring in the Full and Extended tests follows the same rules as those for in-person testing. Given these selection rules, the proportion of scored AMs in a particular life span interval may differ from that typically retrieved. Thus, when presenting aggregate measures of content (i.e., those within a given subject group) not restricted to a particular life period, content values across dating bins are weighted according to the applicable AM temporal distribution.

During online data collection, participants are encouraged to complete the Full test (or Extended test, if opting initially for the Mini test). This is accomplished by pre- and post-test advertisement for the opportunity to explore an interactive summary report of one’s results with the ability to make direct comparisons with results from specified age ranges, solely after completion of the Full (or Extended) test. We stress that, despite their variety (e.g., in duration), all test types provided subjects with the same instruction on AM classification, cueing, dating, and scoring (which were also identical to in-person testing).

When conducting analysis on an individual level, we assumed that a unique test ID corresponds to a unique user. By and large this assumption should be valid; however, it is likely that a subset of users took the test multiple times and were assigned distinct user IDs on each occasion. Moreover, undertaking the Mini or Full version of CRAM following completion of the Atomic format was not coded in the same way as extending the Mini test; in particular, those who opted for the Atomic test and subsequently decided to undergo longer testing were assigned a new user ID. We emphasize that those users who were interrupted or opted to take a break mid-test, however, maintained a single user ID. Likewise, users who opted to extend the Mini test were categorized as taking the Extended test and assigned one unique user ID. As data collected prior to and subsequent to the decision to extend the test was equivalent (data not shown) it appears that practice and in-depth knowledge of dating and scoring procedures did not influence retrieval as measured by CRAM. Unless indicated otherwise, data were collapsed across testing conditions and test types.

As CRAM is freely accessible online⁵ and indexed by popular search engines, data are continuously collected from Internet-browsing individuals. To supplement these unsolicited data, additional individuals were actively recruited from the undergraduate population of George Mason University (GMU), from GMU staff and faculty, and from the local community, obtaining in all cases informed consent. With the exception of undergraduates, recruited subjects were given the choice to complete testing locally at GMU with a researcher present or remotely over the Internet. Recruited undergraduates (ages 18–36 years old) invariably completed the study for course credit and took the test under experimenter supervision; these data from recruited students have been reported previously (Gardner et al., 2012) and included here to best estimate AM content in relatively young subjects. However, the amount of data collected from this age range was substantially augmented by the current approach (exclusively through online testing) almost tripling the previous sample of scored AMs (when pooled together) from these younger subjects. No identifiable personal data were stored. All recruitment and testing procedures were approved by the GMU institutional review board.

In total, 17,482 AMs were dated from 2,561 unique test IDs (Mean Age = 34 years old, SD = 14, range: 18–78 years old; 67% female; 81% native English speakers). A subset of subjects (n = 640) dated AMs, but did not score their content, and were restricted to memory dating analysis. In addition, content measures from 560 Internet-collected AMs (spanning 192 unique test IDs) were inadvertently overwritten prior to back-up, and thus these AMs were also restricted to dating analysis (complete content measures from a subset of AMs from four of these 192 subjects were retained, however, and included in content analysis). After accounting for these events, a total of 6,492 AMs were scored for content (76% scored online) from 1,733 subjects. Fifty-two percent of Internet-scored AMs were collected from the Full test, 20% from the Atomic test, 16% from the Mini test, and 12% from the Extended test; subject age was equivalent across test types (p > 0.10).

Data were further inspected to identify data entry errors or otherwise lazy and inauthentic reporting (e.g., see Gardner et al., 2012). Positive cases were removed from analysis. For example, AMs were excluded if a subject reported an identical number of elements for each of the eight features. In addition, all AMs were removed from seventeen participants whose scoring across the majority of their AMs reflected this pattern (201 AMs in total). Memories were also excluded from two participants who reported either 1 or 11 elements in each feature category across all scored AMs (14 AMs), and from two subjects who reported unique scores for each feature but identically scored all memories (15 AMs). Altogether, these exclusions totaled 232 scored AMs.

Subsequently, extreme total content values were identified as those greater than three times the Inter-Quartile-Range (IQR) above the 75th percentile within a given age range. Data meeting this criterion were considered outliers and excluded from analysis. This procedure was performed separately for AMs collected from each of the following age ranges: 18–25; 26–35; 36–45; 46–55; 56–65; 66–78 years of age. The outlier threshold ranged from 72 to 98 elements per memory depending on the age group (18–25 years old: 72 elements; 26–35 years old: 79 elements; 36–45 years old: 91 elements; 46–55 years old: 98 elements; 56–65 years old: 94 elements; 66–78 years old: 80 elements). This step resulted in the removal of 223 AMs: 78 AMs were excluded from 18 to 25 year old subjects (3.0% of the total within this age range), 52 AMs (3.5%) from 26 to 35 year old subjects, 39 AMs (4.0%) from 36 to 45 year old subjects, 33 AMs (5.1%) from 46 to 55 year old subjects, 16 AMs (4.5%) from 56 to 65 year old subjects, and 5 AMs (2.1%) from 66 to 78 year old subjects. Table 1 provides a summary of the number of scored AMs ultimately retained for data analysis across test types and participant groups.

Analyses were primarily conducted using memory as the unit of observation. Findings were corroborated using an individual subject level approach (see Results). Binary logistic regression was run to assess the effects of participant groups on AM retrieval probabilities across life periods (i.e., Recent and Remote). Bivariate regression was performed to evaluate correlation between measures of AM recall and participant age, and inter-feature relationships. ANOVA was performed to evaluate changes in subject demographics across test types and AM content across participant groups and life periods. Results were corroborated by using a generalized estimating equation approach (Davis, 2002); all conclusions were equivalent with those reported using a general linear model. Where applicable, for robustness analyses (i.e., analysis across genders, native languages, and testing conditions), subject age, and/or cuing and scoring order were assigned as covariates to control for variation in the outcome variable explained by these sources. Chi-square analysis with Yates correction was run to assess group differences in the proportion of AMs assigned to multiple life span bins. Cohen’s d was calculated for each comparison to estimate effect size. Statistical significance was interpreted using the criterion of p < 0.05. False discovery rate correction was applied to multiple comparisons (Benjamini and Hochberg, 1995). Statistical analyses were performed using SPSS (IBM), Excel (Microsoft), and R (Dalgaard, 2008).

Autobiographical memory retrieval probabilities were analyzed against the age of the subject at the time of the recalled episode among various age groups (Figure 1). We observed a large proportion of AMs recalling recent events, which declined steeply with time from the present moment (retention interval). We also found a relative increase in the number of AMs dated to adolescence through young adulthood (the bump), and a relative absence of AMs from early childhood (childhood amnesia).

Moreover, these characteristics of the temporal distribution of AMs produced by CRAM changed with participant age. Retention of Recent AMs was strongest in younger subjects (18–25 years old) and decreased with increasing participant age (p < 0.001). For example, AMs dated to the most recent 10 years of life comprised ∼77% of all AMs cued from 18 to 25 year old subjects, ∼42% of those cued from 26 to 45 year old subjects, and ∼19% of those cued from subjects older than 45 years (see Table 2). Furthermore, while absent in the two youngest age groups, the reminiscence bump emerged in subjects in their mid-to-late 20s, and was most evident in subjects older than 45 years of age (Figure 1). The peak of the bump corresponded to the years between ages 8 and 22, depending on the age group (Figure 1). Age ranges presented in each panel in Figure 1 reflect the changing clarity of the reminiscence bump quantified as the ratio of the peak retrieval probability across life span bins (excluding those associated with the retention interval) to the subsequent minimal retrieval probability. On average, this ratio is undefined in 18–25 year olds (due to the lack of a minimum, reflecting the absence of a bump), greater than one in subjects 26–45 years old, and greater than two in adults older than 45. Childhood amnesia was observed in younger subjects (18–45 years old) as demonstrated by a notable drop in AMs dated to the first 10th of the life span (less than 2%). Its apparent absence in older subjects (i.e., 46–78 years old; Figure 1) is likely an artifact of our methodology. Specifically, the first tenth of life of an older individual extends beyond the relatively narrow temporal period associated with childhood amnesia, and thus limits our ability to isolate and analyze memory for these very early life events.

A total of 6,037 scored AMs were analyzed for content (Table 1; see Section “Materials and Methods” for data screening procedures). On average, subjects reported ∼22 elements (SD = 14) per AM. We found a mild yet significant positive relationship between total reported content and participant age (r = 0.11, p < 0.001). To further investigate this finding, memories were separated into six age groups (see Figure 2A). Upon comparison to the youngest age group (18–25 years old: Mean Total Content = 21.3, SD = 12.1), adults older than 45 years old reported a greater number of remembered details (46–55 years old: M = 24.8, SD = 16.3, d = 0.24; 56–65 years old: M = 25.6, SD = 16.1, d = 0.30; 66–78 years old: M = 24.3, SD = 16.8, d = 0.20); in contrast, reports of content from individuals younger than 46 years old (26–35 years old: M = 20.7, SD = 14.0; 36–45 years old: M = 22.6, SD = 16.4) were found to be comparable to those collected from the youngest group (d = 0.05, and d = 0.09, respectively). Collectively, these findings suggest a non-linear effect of age on total reported content; the age-related increase in the amount of detail reported in typical autobiographical recollection emerges most noticeably in subjects in their mid-to-late 40s and persists into old age (i.e., late 70s; Figure 2A). Given this pattern, to simply describe the magnitude of these effects, AMs were divided into two groups: those collected from subjects 45 years old or younger (Mean Age = 27, SD = 8, range: 18–45 years old; 72% female, 80% native English speakers) and those from subjects 46 years old or older (Mean Age = 57, SD = 8, range: 46–78 years old; 65% female, 89% native English speakers; Figure 2B). This grouping revealed a ∼16% increase in the number of reported details for a given AM in older compared with younger subjects (∼25 vs. ∼21; p < 0.001, d = 0.23; Figure 2B; Table 2). While reported content was quite variable from memory to memory, the coefficient of variation was similar across all ages (∼0.6–0.7; Table 2). The distributions of total content underscore the moderate effect of age on reported details (Figure 2C). Half of all scored memories in younger subjects were comprised of ∼16 or fewer elements; half of all scored AMs in older subjects contained more than ∼20 elements.

Older subjects reported significantly more content than younger subjects for memories of all life periods, both when comparing equivalent decades of life (Figure 3), and when comparing relative life periods (e.g., Recent: the most recent 10 years, and Remote: > 10 years from the present; see Materials and Methods; Figure 3 Inset; Table 2). The numerical results of content analysis of Recent and Remote AMs only marginally fluctuated depending on how these temporal intervals were defined (e.g., restricting Recent AMs to the most recent 5 years; restricting Remote AMs to the first decade of life). In addition, all reported conclusions remained unchanged and did not depend on the particular grouping applied (data not shown).

Older subjects reported ∼24 elements from AMs dated to the first two decades of life compared with ∼20 elements in younger subjects (p < 0.001; d = 0.27). Likewise, Recent AMs from older subjects were comprised of ∼27 elements compared with ∼21 elements in those from younger individuals (Figure 3 Inset; p < 0.001; d = 0.38). Total content declined with the age of the episode among all age groups. For example, Remote AMs were comprised of significantly fewer details compared with Recent AMs (p < 0.01; Table 2; Figure 3). These results confirm those previously reported in college-aged subjects (Gardner et al., 2012) and extend these findings to older adults.

A relatively high proportion of AM content (∼46%; ∼10.1 elements) was associated with Places, Things, and People. In contrast, Times, Contexts, and Episodes were less represented, together comprising just ∼29% (∼6.4 elements). Details and Feelings were close to the average at ∼13% and ∼12%, respectively (Figure 4A). Largely in line with these distributions, the proportion of AMs containing at least one element of a particular feature was highest for Places (98%) followed by People and Feelings (92% each), Things (89%), Details (84%), Times (83%), Contexts (78%), and Episodes (63%). Feature variation was higher than that observed for total content, but equivalent across age groups (Mean Feature CV = 1.0). All these findings uphold those previously reported in younger subjects (Gardner et al., 2012).

Adding to this research, we found that older adults reported a greater number of elements among all features. Episodes and Things showed the most prominent age-related content increase (∼30% and 27%, respectively; p < 0.001), while content associated with People and Times remained close to that observed in young subjects (p > 0.10; Figures 4A,B). These data collectively indicate that the age-related increase in total content is not uniformly distributed across features (in terms of either absolute or relative value). In particular, Things (24%), Episodes (16%), and Details (15%) contributed more substantially to the age-related increase observed in total content (Figure 4C), whereas contributions from People (4%) and Times (3%) were definitively smaller. The proportion of AMs containing at least one element from a given feature was significantly higher in older compared with younger subjects for Feelings, Things, Details, Contexts, and Episodes (∼5% higher on average; p < 0.001) but not for People, Places, and Times (p > 0.10). These feature distributions were consistent between Remote and Recent AMs in both younger and older subjects (Supplementary Figure S1); notably, the feature Times appeared to be least resilient to temporal decay of content among all subjects.

Content correlation analysis showed positive relationships among all features (mean Pearson r = 0.37). Moreover, these relationships were similar across age groups (younger: r = 0.36; older: r= 0.40) as well as between Remote and Recent intervals (Remote: r = 0.39; Recent: r= 0.37). To further evaluate feature dependence, the correlation between each feature and all other content was computed. The average of these values across all features was equivalent between age groups (younger: r = 0.53; older: r = 0.57; Supplementary Figure S2A) and across life periods (not shown). However, the feature People exhibited a comparatively mild relationship (r = 0.37; 32% less than the average: Supplementary Figure S2B) across all conditions, suggesting that it is a relatively independent component of recall. The inter-feature relationships found here confirm those previously reported among younger subjects (Gardner et al., 2012) and extend these findings to individuals distributed across the life span.

This work provides numerical description of AM retrieval probabilities and reported content associated with distinct life periods. Combining these two measurements, we can estimate the relative distribution of retrieved content across temporal intervals and age groups (Figure 5). Specifically, given the number of elements typically reported for a single retrieved AM, for each of the various age groups (as outlined in Figure 2A and plotted in Figure 5 by mean age), we computed a probabilistic content distribution among life periods. For example, when experiencing AM, a typical 70 year old, on average, reports a similar content amount from his or her middle teenage years compared with that from the last few years of his or her life. In contrast, content from events dated to the most recent 2 years of the life of a 21 year old is ∼5 times more represented in memory than that associated with events from his or her middle teenage years. Assuming that the frequency of AM recollection is stable across age groups (e.g., see Gardner and Ascoli, 2015), we can further estimate the relative probability that a particular recalled element is associated with a certain age group and life period. For example, the likelihood for a 30 year old to retrieve content from his or her early 20s (∼1.3%) is equivalent to that for a 60 year old to retrieve content from his or her early thirties.

Autobiographical memories found within the reminiscence bump may have distinct recall characteristics. For example, AMs that compose the bump may be comparatively rich with recalled detail. Such a finding would explain, at least in part, why this life period plays a particularly prominent role in subjective experience. We evaluated reported content from AMs within and beyond life periods associated with the bump in older subjects (46–78 years old) for whom this phenomenon was most pronounced. In this age range, we observed a relatively high AM retrieval probability (excluding temporal bins associated with the retention interval) from ages 11–20 and a relatively low probability from ages 31–40 (see Figure 1). However, measures of total content from these life periods did not significantly differ (M = 24.90 and M = 25.28, respectively; p > 0.10, d = 0.02; Figure 3). The composition of memories dated to these life periods was also quite stable (p > 0.10), with any given feature showing a deviation of ∼2% or less (Mean deviation: ∼1%).

All the main findings of this work are robust to gender, native language, and changes in experimental procedures (see Supplementary Tables S1 and S2). For example, total reported content moderately increased with age both for females and males, as well as for native and non-native English speakers (see Supplementary Material). However, several minor quantitative distinctions in AM recall were observed between males and females, between native and non-native English speakers, and among testing conditions. A full review of these results is included as Supplementary Material.

As prior reports suggest that different individuals have distinct recollection experiences (Rubin et al., 2004), the main conclusions of this work were evaluated using a subject-level analysis. Given the relatively large amount of variability in content scores from memory to memory (see Table 2), the analysis was restricted to those subjects who scored five or more memories (18–45 years old: n = 496; 46–78 years old: n = 118). This restriction permitted measures that more closely represent typical recollection from a given individual. In addition, for inclusion in comparisons across Recent and Remote intervals, subjects were required to have an AM scored in each temporal period (18–45 years old: n = 420; 46–78 years old: n = 69). Content measures reported independent of temporal periods were weighted according to the temporal distribution of AM retrieval computed separately for each subject on the basis of his or her dated AMs.

Age effects on the temporal distribution of AMs and AM content are confirmed by subject-level analysis. In particular, the proportion of AMs from Recent life intervals was relatively high in young subjects (18–25 years old: ∼77%) and decreased with increasing age (26–45 years old: ∼40%; 46–78 years old: ∼17%). Moreover, older adults (46–78 years old) reported ∼15% more elements than younger (18–45 years old) subjects (Mean ± SD: 25 ± 13 compared with 21 ± 10 details; p < 0.01), replicating age differences found using memory as the observational unit. Remote memories contained fewer elements (18 ± 11) than Recent memories (23 ± 13; p < 0.001) across all subjects, and older compared with younger subjects reported more elements from Recent (Older subjects: 29 ± 18; Younger subjects: 22 ± 11; p < 0.001) and Remote events (Older subjects: 23 ± 13; Younger subjects: 18 ± 10; p < 0.001). In addition, an age-related increase in total reported content was most drastic for Episodes (∼31%) and Things (∼30%), whereas considerably less dramatic for Times ( < 1%) and People (∼4%), closely resembling findings using memory as the unit of analysis. Likewise, total content scores from older subjects did not discriminate between bump and non-bump memories (e.g., comparing content from AMs dated to when subjects were 11–20 years old with those dated to when subjects were 31–40 years old; p > 0.10; data not shown).