Learning Strategy Differentially Impacts Memory Connections in Children And Adults(1)

Abstract 

Even once children can accurately remember their experiences, they nevertheless struggle to use those memories in flexible new ways—as in when drawing inferences. However, it remains an open question as to whether the developmental differences observed during both memory formation and inference itself represent a fundamental limitation on children’s learning mechanisms or rather their deployment of suboptimal strategy. Here, 7–9-year-old children (N = 154) and young adults (N = 130) first formed strong memories for initial (AB) associations and then engaged in one of three learning strategies as they viewed overlapping (BC) pairs. We found that being told to integrate—combine ABC during learning—both significantly improved children’s ability to explicitly relate the indirectly associated A and C items during inference and protected the underlying pair memories from forgetting. However, this finding contrasted with implicit evidence for memory-to-memory connections: Adults and children both formed A-C links prior to any knowledge of an inference test—yet for children, such links were most apparent when they were told to simply encode BC, not integrate. Moreover, the accessibility of such implicit links differed between children and adults, with adults using them to make explicit inferences but children only doing so for well-established direct AB pairs. These results suggest that while a lack of integration strategy may explain a large share of the developmental differences in explicit inference, children and adults nevertheless differ in both the circumstances under which they connect interrelated memories and their ability to later leverage those links to inform flexible behaviors. 

Cistanche extract powder

Click here to view  Cistanche Improve Memory products

【Ask for more】 Email:cindy.xue@wecistanche.com /  Whats App:  0086 18599088692 /  Wechat:  18599088692

KEYWORDS 

development, explicit memory, implicit memory, inferential reasoning, memory integration, priming,Cistanche deserticola

Cistanche powder

1 INTRODUCTION

Memories can be used not only to reflect upon the past but also to guide present and future behavior (Mack et al., 2018; Schlichting & Preston, 2015; Zeithamova et al., 2012). Indeed, we often draw upon memories of specific episodes to solve unanticipated problems (Carpenter et al., 2021; Giovanello et al., 2009; Zalesak & Heckers, 2009) or imagine never-experienced scenarios (Addis et al., 2011; Schacter et al., 2012). This ability to generate new information by integrating across multiple events is a key function of memory (Sheldon & Levine, 2016; Zeidman & Maguire, 2016)—and yet, one that might emerge relatively late in development (Bauer et al., 2015; Coughlin et al., 2014; DeMaster et al., 2015; Ghetti & Coughlin, 2018; Schlichting et al., 2017, 2022; Shing et al., 2019; Varga & Bauer, 2013). Specifically, past work has shown that children struggle to use their memories flexibly, even despite remembering the underlying experiences (Bauer & San Souci, 2010; Schlichting et al., 2017, 2022). Mechanistically, it has been suggested that such difficulty might lie in when memory-to-memory connections are formed: While adults can integrate during learning (i.e., engage in ‘integrative encoding’; Schlichting & Preston, 2016; Schlichting et al., 2014, 2015; Shohamy &Wagner, 2008; Varga & Bauer, 2017a; Varga & Manns, 2021; Zeithamova & Preston, 2010; Zeithamova, Dominick & Preston, 2012), children and adolescents might instead wait for an explicit prompt at test (Bauer et al., 2015, 2020a; Varga & Bauer, 2013). These tendencies would yield different memory structures with distinct associated demands for a later test: While adults may already boast the connections needed to support the new decision, children instead would need to make such connections during the judgment itself—taxing young learners with additional operations and ultimately impeding performance.

Cistanche deserticola experiment

Despite growing research on this topic, there remain at least two (non-mutually exclusive) possible explanations as to why children do not integrate until prompted. First, it may be the case that integrative encoding—a complex, multi-step process (Bauer & Varga, 2017; Zeithamova, Schlichting, et al., 2012), requires first reactivation of related memories during learning followed by integration upon detecting associative novelty (Schlichting et al., 2014; van Kesteren et al., 2020; Zeithamova & Preston, 2017; Zeithamova, Dominick, et al., 2012)— is simply neurocognitively out of reach for children. Children would fail to integrate if the cognitive mechanisms supporting any of these steps were immature—and indeed, past work has shown that children under 10 years of age do not engage in even the first of these steps (reactivation; Miller-Goldwater et al., 2021; Schlichting et al., 2022). However, a second possibility is that adults, but not children, deploy a top-down integration ‘strategy,’ which could widen the developmental performance gap. Research in adults suggests that although integrative encoding can be engaged in the absence of awareness (Shohamy & Wagner, 2008), both awareness (Varga & Bauer, 2017b) and instruction (Burton et al., 2017) increase their likelihood. Because adults may be better poised to detect the relationships between experiences and/or anticipate a wider array of memory tests than children, they may adopt a more advantageous learning strategy accordingly (Shing et al., 2008, 2010). If the lack of an integration strategy is the main source of developmental difference, it would follow that instructing children to integrate during learning would enhance performance. However, if instead, children are not yet able to integrate, then encouraging them to do so would not help—and might even hurt—their performance. We assess these possibilities here.

We additionally reasoned that many explicit tests of memory and inference put children at a disadvantage—due not necessarily to the memory content itself, but rather to demands (e.g., control, interference resolution, selection) related to test format. We aimed to disambiguate these two possible sources of developmental change. As such, we adopted a priming task as an implicit—and relatively purer— measure of memory-to-memory connections (Davis et al., 2021). Past work using similar approaches has shown that priming (i.e., facilitated processing of a target item when preceded by its associate) is sensitive to even arbitrary, experiment-defined pairings (McKoon & Ratcliff, 1979); it is also present even in young children and relatively stable over-development (McCauley et al., 1976; McFarland & Kellas,1975; see also Hasher & Zacks, 1979; Parkin & Street, 1988). Coupled with typical explicit assessments, we were therefore poised to ask whether development is attributable mainly to a differential tendency to form associative connections (which would be evident in both implicit and explicit tests), or rather differential ability to engage top-down influences at retrieval (only explicit).

Main Chemical Constituents of Cistanche deserticola

We had children and young adults learn overlapping associations (AB, BC) that provided an opportunity for integration (ABC). We tested 7–9-year-old children as we anticipated that while they would be capable of doing the task, they would nevertheless accomplish it differently than adults (Schlichting et al., 2017, 2022; Shing et al., 2019; Wilson & Bauer, 2021). Our key manipulation was in the instructions participants received prior to viewing BC pairs: one-third of our participants were told each to retrieve (i.e., recall the related AB pair and ignore BC), encode (focus on the current BC), or integrate (recall A and combine it with BC; an instruction shown to increase integration among adults; Burton et al., 2017; Richter et al., 2015). Integration was indexed as a connection between A and C items, which we quantified first in implicit priming and then explicit inference tests. We hypothesized that in contrast to adults, children would both (1) not exhibit A-C connections in the priming task because it occurred prior to a prompt; and (2) perform best on explicit inference when encouraged to simply encode BC due to their reduced capacity for integrative encoding. We additionally anticipated that children instructed to retrieve might be unable to ignore BC, and therefore have paradoxically better memory for the (irrelevant) BC pairs than adults.

Cistanche supplement near me-Improving Memory

2 METHOD

Our hypotheses, sample size, and analysis approach were preregistered there (https://osf.io/fmv3w/), with deviations as noted.

2.1 Participants

Results are from 130 adults (96 female, 34 male; age range = 25.02– 35.95 years [y], mean = 29.17, standard deviation [SD] = 2.99) and 154 children (80 female, 74 male; age range = 7.01–9.93 y, mean = 8.51, SD = 0.85) who met our preregistered inclusion criteria (see Supplementary Information for full sample details). Fifty-one additional participants were tested and excluded for reasons related to their performance: failure to perform the instructed encoding strategy (perhaps due to poor instruction comprehension, described below; 18 children, seven adults); and subthreshold memory for the initial (AB) pairs (defined as <80% correct on the last test round of the initial learning phase; 22 children; four adults). Our reason for requiring near-perfect memory for AB pairs was to ensure that the results are based on our manipulation during the overlapping (BC) exposure and not due to inadequate initial memory. Comparing participants who were ultimately included versus those who were excluded for performance-related reasons revealed no significant differences in working memory in either age group (two-sample t-tests; children: t(48.72) = 1.66, p = 0.102; adults: t(11.06) = 1.04, p = 0.319).

Participants were randomly assigned to an instruction condition (retrieve, encode, or integrate), and conditions did not differ in terms of age (children: F(2,151) = 0.729, p = 0.484; adults: F(2,127) = 0.162, p = 0.850) or working memory (Supplementary Methods; children: F(2,142) = 2.696, p = 0.071; adults: F(2,125) = 0.326, p = 0.723).

2.2 Stimuli

Stimuli were pictures of 90 common objects selected as being likely familiar to 7-year-old children (based on age-of-acquisition norms; Kuperman et al., 2012). Objects were organized into 30 ABC ‘triads’. Each triad consisted of a fixed set of A, B, and C objects of low semantic similarity as determined using WordNet::Similarity. Of the 30 triads, six were ‘catch’ and 24 were ‘experimental’ triads. Catch triads were included purely to ensure participants were following instructions and were excluded from all analyses. Assignment of stimuli to A, B, and C positions within a triad was counterbalanced across participants such that each object occurred equally often in each position.

2.3 Procedure

Participants provided informed consent (adults) or verbal assent (children); parents/guardians also provided permission for children. Procedures were approved by the ethics committee at our institution, and participants were compensated for their time at a rate of $10/hour.

Sessions were run remotely via video conferencing (Zoom), and the task was presented to participants on their personal computers using Inquisit (programmed in Inquisit 5; https://www.millisecond.com). Each session lasted on average 1.5–2 h. Participants viewed ABC triads as overlapping AB and BC pairs in an associative inference task (Preston et al., 2004) with four phases: Initial (AB) learning, overlapping (BC) exposure, preference (priming) task, and final (AC inference and direct) test (Figure 1). Since we manipulated task instructions and experiences were otherwise identical across participants, experimenters were hypothesis-blind and only un-blinded after data collection. Task instructions and practice trials were given to participants immediately before each phase, as described below. For additional procedural details, see Supplementary Methods

2.3.1 Initial (AB) learning

During this phase, participants established near-perfect memory for all 30 AB pairs across three study-test cycles. Prior to beginning AB learning, participants were told that they would see two objects on the screen and that they should create a story to help them remember the pair. They were also informed as to how they would be tested on their memory for the AB pairs and encouraged to think back to the stories they made up during learning to help them remember. Participants completed a practice that was identical to the real AB learning, but with fewer pairs and stimuli that were separate from those used in the main experiment.

FIGURE 1 Task schematic depicting four experimental phases (colored boxes; top and bottom) and their order (timeline; middle). During AB learning (top left, yellow), participants studied AB associations (e.g., bed-watermelon) across three study-test cycles. Then, during BC exposure (top right, navy), participants saw a single presentation of overlapping associations (e.g., watermelon-turtle) with which they performed one of three tasks (teal, purple, pink)—retrieve, encode, or integrate—and then took a memory test. Next, participants completed a priming task designed to measure whether connections existed between A and C items in memory (bottom left, green). Predictions are schematized for an intact (left) and rearranged (right) pairing, where the presentation of the first object (e.g., lettuce) would facilitate the processing of the second (lamp) if the objects were associated in memory (intact; denoted by green connecting line and check mark) and slowed if they were not (egg and pacifier are a rearranged pair; red line and x). Lastly, participants took a final test (bottom right, red) that included both inferential (left) and direct (right) associations.

During study trials, participants viewed an AB pair (A on the left, B on the right) and were asked to create a story relating the two objects. For most trials (experimental; 5 s stimulus with 1 s interstimulus interval [ISI]), participants did not make any overt response but internally rehearsed their story; however, for some trials (catch, first repetition only; 14 s stimulus with 1 s ISI), participants were cued with the text ‘Tell me your story’ appearing on the screen to say their story aloud. Catch trial stories were transcribed by the experimenter and subsequently scored by the first author (ZA) to ensure compliance (Supplementary Methods).

After each study, participants completed a self-paced three alternative-forced-choice (3AFC) test for all 30 AB pairs. Participants were cued with A and were asked to select the associated B. Incorrect options (foils) were familiar B stimuli from other triads. Moreover, experimental triads were never foiled with catch triads or vice versa, such that participants would not be able to use memory for the nature of their response at encoding (i.e., whether the story was produced aloud or rehearsed internally) to influence their selection. Once the participants completed the three study-test blocks, they took a 5-min break during which time they worked on an unrelated puzzle.

2.3.2 Overlapping (BC) exposure

Having formed strong memories for the initial AB pairs, participants were then shown one repetition of each BC pair (i.e., single-shot learning). Prior to beginning this phase, all participants were informed that they would see new pairs but that one of the objects would be old because it was shown in a pair they saw in the previous learning phase. Participants were also told they would be tested on their memory in the same way as before.

On each trial, participants were presented with a BC pair (B on the left; experimental trials were 5 s stimulus, 1 s ISI) and asked to create a story to help their memory. Participants were given different instructions about the kind of story they should create, which yielded the retrieve, encode and integrate conditions (Richter et al., 2015; in our study, this was an across-participants manipulation). Participants in the retrieve condition were instructed to ignore the current BC pair and think back to their related AB pair story; in the encoding condition, to create a story relating the two new BC objects; and in the integrated condition, to recall the associated A object and create a story relating all three (A, B, C) objects. Encode and integrate strategies were meant to mimic separate encoding of AB and BC for later recombination (somewhat akin to what we posited as the naturalistic tendency of children) and ABC integration (the tendency of adults), respectively. The retrieve condition was additionally included to assess whether children form associations among to-be-ignored stimuli, yielding the possibility that children might outperform adults in BC memory. We assessed participants’ compliance with these instructions through the inclusion of catch trials, in which they spoke their stories aloud (as in AB learning; 14 s stimulus, 1 s ISI). Participants had an opportunity to practice the learning task with separate pairs prior to beginning the real BC learning phase and were provided with corrective feedback if they incorporated the wrong items into their practice stories.

After BC exposure, participants were tested on their memory for all BC pairs using a self-paced 3AFC test that was similar in format to the AB test. B objects served as prompts, and the correct C objects were presented among foils from other triads of the same type (experimental or catch).

2.3.3 Preference (priming) task

Next, participants completed a preference task designed as an indirect measure of associative memory. Participants were told that this was a new game where we were interested in knowing what kinds of things they like and dislike; no reference was made to the relation between this task and the other task phases. Experimenters reassured participants that there were no right or wrong answers since everyone has different opinions. Participants were simply asked to try their best to think about each object before answering throughout the game and to use both like and dislike options. Before beginning the real task, participants completed a short practice task with separate stimuli.

In each trial, participants viewed one A, B, or C stimulus. The stimulus period was 1.5 s, during which time participants viewed a stimulus (0.5 s) and made a preference judgment (i.e., indicate whether they liked or disliked the object; additional 1 s response window). Trials were jittered such that stimulus onsets occurred at inter-trial intervals [ITIs] of 3 s, 4.5 s, and 6 s (or 1.5 s, 3 s and 4.5 s ISI) on 40%, 40%, and 20% of trials, respectively (timing modeled after Turk-Browne et al., 2012).

Our key question was how participants’ response times (RTs) were influenced by the particular object sequencing. Specifically, objects were preceded by other objects from either the same or a different triad (triads were assigned to half each ‘intact’ and ‘rearranged’ conditions, respectively). We predicted that processing would be facilitated in the former (and therefore, yield faster RTs) relative to the latter case if participants had memories for the pairs. We embedded in our sequence the indirect AC (first 1/3 of the sequence) as well as direct AB and BC (second 2/3 of the sequences; intermixed) pairs, enabling us to ultimately derive a separate priming measure for each pair type (AC, AB, BC). The task was divided into three blocks, and participants had the opportunity to take a short rest between blocks if they wished.

2.3.4 Final (inference and direct) test

Next, participants were informed about the overlapping nature of the AB and BC pairs and were instructed that they should infer a relationship between A and C objects that had been associated with the same B object. That is, all participants were cued to integrate. The experimenter walked participants through an example using stimuli from the instructions, but there was no separate practice task. Following these instructions, participants completed the final 3AFC tests over first AC inference (A served as the prompt) and then direct AB and BC (intermixed, tested in the same way as in the initial tests) associations.

2.4 Comprehension of instructions

Participants stated their stories aloud on catch trials during AB and BC learning, which allowed us to determine their comprehension of and compliance with the task instructions. Specifically, after the session we scored whether participants’ stories were correct (i.e., aligned with the instructions) or incorrect, and excluded those who did not get at least 3/6 stories correct for each phase. As noted in Participants, this criterion ultimately led to the exclusion of 18 children (five retrieves; 13 integrate) and seven adults (two retrieves; two encode; three integrate). We assessed whether these exclusions varied significantly by age group, condition, or their interaction using a general linear model with a binomial linking function. We found trend-level evidence for the interaction (age group × condition: χ2(2) = 5.673, p = 0.059), with children being more likely to be excluded than adults in the integrate (z = 1.960, p = 0.050) but not the other two conditions (both |z| < 0.936, both p > 0.349). This result suggests that the integration condition was especially difficult for children. However, we underscore that after these exclusions, all participants ultimately included in our analyses both understood and complied with the task instructions.

2.5 Statistical analyses

Our statistical analysis approach is described in the Supplementary Methods. Briefly, we analyzed data in R (Team, 2018) using mixed-effects models (Bates et al., 2015). For accuracy and RT on explicit memory and inference tests, we assessed both the main effects and the interaction of instruction condition and age group. Predictors for initial AB learning were repetition × age group interaction and main effects since the instruction manipulation had not yet been introduced. For the implicit measure of memory, we modeled RTs on the priming task within each instruction condition separately as a function of the interaction of age group (child vs. adult) and sequence type (intact vs. rearranged).

3 RESULTS

3.1 Robust AB learning in all instruction conditions

AB memory improved across learning (main effect of repetition; z = 14.22, p < 0.001), becoming near-perfect by the third and final repetition for both age groups (children: mean = 97%, SEM = 0.004; adults: mean = 99%, SEM = 0.002; Figure S1). Children’s performance was overall lower than adults’ (z = 7.04, p < 0.001) and showed a statistical trend for steeper learning slopes (marginal condition × repetition interaction; χ2(1) = 3.242, p = 0.072). Importantly, considering performance at the end of AB learning within age revealed no differences as a function of instruction condition (both χ2(2) < 1.79, p > 0.408), underscoring that our groups were matched in overall memory ability before introducing our manipulation. In subsequent analyses, we consider only new learning and integration associated with those AB pairs that were initially learned—that is, correct on the final learning repetition—unless otherwise noted.

3.2 Explicit memory and inference performance underscores rigidity in children

Next, we assessed whether being instructed to retrieve (AB), encode (BC), or integrate (ABC) during BC exposure differently impacted memory for the single-shot BC pairs and/or later AC inference across age groups.

With respect to BC learning (i.e., performance on the memory test immediately following BC exposure; Figure 2A), the impact of instruction condition on accuracy (Figure 2B) significantly differed between children and adults (instruction condition × age group interaction: χ2(2) = 12.503, p = 0.002). Adults performed significantly better than children under encode and integrate (both z > 2.569, p < 0.011) instructions. The retrieve condition asked participants to ignore the BC pairs of interest here. In this scenario, adults performed no better than children (z = 1.575, p = 0.115). Moreover, BC performance was numerically highest following instructions to encode in both age groups: Adults showed a significant advantage of encode over both retrieve (z = 9.278, p < 0.0001) and integrate (z = 4.500, p < 0.0001; integrate was also significantly better than retrieve: z = 5.222, p < 0.0001). For children, the advantage for encoding was significant relative only to retrieve (z = 5.762, p < 0.0001; not different for integrate, z = 1.263, p = 0.207). A complementary pattern was also observed in RTs, with both children and adults being significantly faster to make accurate BC decisions following encode than integrate instructions (adults: z = 3.669, p = 0.0002; children: z = 2.468, p = 0.014; Supplementary Results and Figure 2C). Overall, these results suggest that being encouraged to focus solely on encoding the current BC experience was best for both children and adults, such that the added task of incorporating the held-out A object (as in integrate) impeded BC memory (albeit significant for children in RT but not accuracy).

We next asked how instructions during BC exposure impacted participants’ ability to draw novel AC inferences (Figure 3A). Importantly, we restricted to triads for which the corresponding AB and BC pairs were initially learned to control for differences in memory for the direct pairs. Adults were more accurate than children in all conditions (Figure 3B; all z > 2.405, p < 0.017). There was also a significant interaction of instruction condition and age group (χ2(2) = 18.119, p = 0.0001) that revealed a strikingly different pattern in children and adults.

Specifically, children were more accurate following instructions to integrate than either encode or retrieve (integrate vs. encode: z = 4.443, p < 0.0001; integrate vs. retrieve: z = 4.858, p < 0.0001), while encode and retrieve were not different from one another (z = 0.537, p = 0.591). Therefore, despite forming strong BC memories (Figure 2B), children following the encoded instructions struggled to make new connections across those memories during explicit AC inference (Figure 3B). In fact, children told to encode BC showed low inference performance that— though significantly above 33.3% chance (CI95% = [0.37, 0.52])—was still no better than those children instructed to ignore BC entirely

Adults, by contrast, were more accurate for both the integrate (z = 5.433, p < 0.0001) and encode (z = 5.844, p < 0.0001) relative to retrieve conditions; integrate and encode were not different from one another (z = 0.498, p = 0.618). In other words, they achieved similar accuracy on AC inferences based upon BC memories formed under either encode (83%) or integrate (82%) instructions. However, adults did show a statistical trend for a difference between these two conditions in terms of speed: They were marginally faster to make correct inferences when told to integrate than encode during learning overall (t(225) = 1.741, p = 0.083; Supplementary Results and Figure 3C). Inferences were significantly speeded for integration relative to encode instructions when accounting for differences in direct BC RT (instruction condition × pair type interaction, omitting the low-accuracy retrieve condition: χ2(1) = 37.775, p < 0.0001; the pattern was such that BC memory was significantly slower, z = −3.596, p = 0.0003, yet inference significantly faster, z = 2.314, p = 0.021, for integrate versus encode; there was no such interaction for children: χ2(1) = 1.755, p = 0.185).

In summary, considering BC memory and AC inference revealed rigidity in children, in that they performed best on the test that most closely matched their BC instructions: Those told to encode achieved the highest BC accuracy, whereas those in the integrated condition excelled at AC inference. Moreover, despite children in the encode condition learning BC well, they nevertheless struggled to apply those memories flexibly to a new, ‘uninstructed’ (i.e., not aligning with exposure instructions) test. (This dissociation also underscores the effectiveness of our manipulation: No two instruction conditions were alike in terms of both BC learning and AC inference performance, suggesting that children were modulating their engagement with the material to yield unique retrieve, encode and integrate profiles.) In contrast, adults were more flexible in that they were able to achieve similar inference accuracy with BC memories formed under encode and integrate instructions—though the former did take marginally more time.

Our child's age range spanning several years (7–9) allowed us to ask whether there was a transition away from memory rigidity during this period of development, which would provide insight into when children become able to use their memories in a flexible manner. Interrogating performance as a function of age (Supplementary Results) revealed a significant age × trial type (BC vs. AC) × instruction condition interaction (χ2(2) = 7.828, p = 0.020) underpinned by significant relationships with accuracy unique to the uninstructed test. That is, children in the encode condition showed age-related gains in AC inference (AC: z = 3.30, p < 0.001; no relationship for BC: z = 0.63, p = 0.529), whereas those in the integrated group showed improvements only on BC memory (BC: z = 2.95, p = 0.003; trend in AC: z = 1.69, p = 0.091; no such relationship for retrieve, both |z| < 1.10, both p > 0.277; Figure 4A). Moreover, we observed an increasing RT cost unique to the encoding group (Figure 4B), with 9-year-olds’ RTs mirroring what we observed in adults (Figure 3C). Overall, this finding is consistent with the emergence of memory flexibility during this period in childhood, perhaps at least partially underpinned by a maturing capacity for test-phase memory restructuring.

FIGURE 2 BC exposure task performance. (a) The task timeline (top) is depicted in Figure 1. Participants viewed BC pairs (left) and performed one of three tasks (thought bubbles). They then took a 3AFC BC test (right), the accuracy and RTs which are depicted in panels B and C. (b) BC accuracy by instruction condition, restricted to pairs for which the corresponding AB was correct during the last learning repetition. (c) RTs on correct BC test trials. For both b and c, points represent participant means. Black circles and confidence intervals represent estimated marginal means and 95% confidence intervals from mixed-effects models. The dashed line represents chance performance. Black significance markers denote pairwise comparisons across instruction conditions within age groups; color-coded markers denote comparisons across age groups, within instruction conditions. * p < 0.05

3.3 Implicit A-C Connections in children and adults prior to explicit inference

We reasoned that above-chance performance on explicit inference might be supported by either (a) separate memories for the individual premise pairs (AB, BC) along with recombination during inference itself; or (b) A–C connections formed during learning. Therefore, we next turned to an indirect assessment that allowed us to ask whether the presence of such connections in memory varied by development stage (child, adult) and/or instruction condition (retrieve, encode, integrate). Of note, this assessment occurred prior to participants being informed about the relationship between AB and BC pairs, and before explicit AC inference. Therefore, any evidence for A–C connections at this point in the experiment would indicate their formation in the absence of overt demand.

Our approach was to assess whether participants were speeded in making decisions about C objects that were preceded or ‘primed’ by their indirectly associated A object (‘intact’ pairings) as compared with an unrelated A (‘rearranged’). (We focus on the A-C portion of the priming task which occurred first, as we saw evidence for memory disruption—primarily in adults—that might explain the lack of priming we observed for direct associations tested later in the task; see Influence of Preference Task on Memory in Supplementary Results.) Our logic was that, if and only if A and C were connected in memory, processing of A would facilitate the processing of the associated C—even on an unrelated preference judgment. We, therefore, compared priming task RTs between intact and rearranged targets to investigate whether evidence for such facilitation existed among children and adults who had received retrieve, encode, and integrate instructions during BC exposure. Of note, here we deviated slightly from our preregistration by including trials irrespective of subsequent AC inference success, as we found developmental differences in the correspondence between these two tasks (see below). Our initial hypothesis was that adults but not children would spontaneously connect A-C items in memory, with developmental differences being significant only in integration. Moreover, we predicted that adults would show sensitivity to instructions, with A-C connections being most evident in integration.

FIGURE 3 Inference performance (explicit measure). (a) After BC exposure (blue in the top timeline; with instruction manipulation, left) and the priming task (green in the timeline), participants completed an explicit inference test (red in the timeline; right). (b) Accuracy on AC inference decisions, considering only AC trials for which the associated AB and BC were both correct during learning (final repetition for AB pairs). (c) RTs on correct trials are also restricted to correct direct pair memory as in panel b. Data are depicted in Figure 2. * p < 0.05; ∼ p < 0.10

FIGURE 4 Performance (y-axes) on BC pairs (solid lines) and AC inferences (dashed lines) as a function of age in years (x-axes) among children. (a) Accuracy and (b) RT. Data are the same as in Figures 2 and 3, but re-plotted by age. For all plots, colored points represent individual participant means for BC memory (filled circles) and AC inference (open circles). Lines and 95% confidence bands are derived from mixed-effects models. Full statistics are provided in the Supplementary Results. * p < 0.05; ∼ p < 0.10

Results were partly consistent with these predictions (Figure 5). Specifically, we found that on average, adults in only the integrate condition (z = 2.262, p = 0.024; and not the other two, both |z| < 0.989, both p > 0.322) exhibited significant priming. However, contrary to our hypothesis, this effect was not unique to adults: Children in the encode (z = 2.375, p = 0.018; but not in integrating, z = 1.320, p = 0.187; or retrieve, z = 0.939, p = 0.348) condition also stored A-C connections in memory. Moreover, we did not observe a significant age group × sequence type (intact vs. rearranged) interaction in any instruction condition (all χ2(1) < 1.140, p > 0.285), suggesting children and adults do not differ from one another in the degree of priming. Together, these results suggest that while children and adults alike connect indirectly related memories during new learning experiences, they might be most apt to do so under different learning strategies.

FIGURE 5 A–C priming task (implicit measure). Overall priming effects for A–C pairs by age group (children, left and adults, right violin in each pair) and instruction condition (colour). Coloured points are individual participant means; black points represent estimated marginal means and 95% confidence intervals from mixed-effects models. * p < 0.05 versus zero