The Flexibility of Working Memory in Drawing on Episodic Long-Term Memory Representations in Serial Recall

Method

Materials and procedure

All experiments were programmed in jsPsych (de Leeuw et al., 2023) using the jspsych-psychophysics plugin (Kuroki, 2021). Figure 1 provides an overview of the general procedure of Experiment 1. It consisted of two main phases: an LTM learning phase and a WM task phase. Stimuli consisted of 320 English nouns chosen randomly for each participant from a set of 1192 nouns. Words were concrete, had a minimum and maximum length of 3 and 7 letters, and had an average frequency of 48.57 (SD = 94.11) according to the recommended Hyperspace Analogue to Language -HAL -criterion (Balota et al., 2007). Out of the 320 words, 40 were used to form 20 LTM word pairs.

Figure 1

As depicted in Figure 1A, the experiment started with the LTM learning phase, which consisted of the presentation of 20 word pairs divided into 4 mini-blocks. Each mini-block sequentially displayed 5 word pairs at the centre of the screen for 3500 ms (e.g., jeans-cat) with an inter-stimulus interval (ISI) of 500 ms. This was followed by a test of each word pair in random order, in which the first word of each pair was presented as a probe, and participants were required to type in the associated word. After participants pressed the “enter” key, feedback was provided. If the response was correct, a message displaying “Correct!” in green appeared immediately. If the response was incorrect, an “X” with the correct word shown in red appeared instead. Feedback lasted 2000 ms. This process was repeated a second time, with a new random order of the word pairs across the mini-blocks. Finally, participants completed a final test for all the 20 word pairs in random order following the same test procedure described earlier. These testing phases interspersed with the study were intended to boost memory through the testing effect (see Sutterer & Awh, 2016 for a similar approach), as testing has shown to improve the retention of information on the long-term (Roediger III & Karpicke, 2006).

The WM task phase consisted of the immediate serial recall of 6 words presented sequentially, each for 1000 ms with an ISI of 500 ms. This phase was divided into 3 blocks, each consisting of 20 trials. Each block represented one of three conditions: The word pairs, the singletons and the new condition. Blocks and thereby the order of conditions were counterbalanced across participants. In the word pairs condition an LTM word pair was presented in succession within the list, retaining the order in which it was originally learnt (e.g. jeans – cat). The position of the word pair was manipulated across trials, so that across the entire experiment, it appeared twice at each possible position within the list (e.g., at position “1,2”, the first word of the pair was presented at serial position 1, and the second associated word at serial position 2). In the singletons condition two single words of two different episodes (i.e., LTM word pairs) were presented in succession within the list (e.g. jeans – rope). As in the word pairs condition, the position in which the words appeared within the list was counterbalanced across trials. In the new condition, only new words were presented. This served as our baseline condition to measure any beneficial effects of episodic LTM to the WM task (see Figure 1B).

After the presentation of a word list, immediate memory for serial order was tested. For half of the trials, immediate memory was tested via a typed recall test, while for the other half, a 12-alternative forced choice (12-AFC) recognition test was employed. The results of the 12-AFC recognition can be found in the supplementary materials. In brief, results are consistent with the findings of the typed recall test that are described below. The tests were inter-mixed within a block. For the typed recall test, participants were required to type each word in a box at the centre of the screen, in the same order as they were presented. Each box provided a cue indicating the word to type (e.g., “Word 1”), and after typing the word and pressing “enter”, a new empty box with the proper following cue appeared (e.g., “Word 2”). Recall requires retrieval of specific and contextual details (Yonelinas, 2002), which makes the test specifically sensitive to detecting benefits of information presented as an intact episode (i.e., in the word pairs condition).

Data Analysis

All analyses were conducted in a Bayesian framework using R, version 4.3.2 (R Core Team, 2023 with the following main R packages: tidyverse (Wickham et al., 2019), brms (Bürkner, 2017), bayestestR (Makowski et al., 2019). Data was analysed on a trial by trial basis, using the number of correct responses out of the total number of responses within a trial as the dependant variable. Correct responses were defined as accurately recalling a word in the position it was originally presented in –strict recall.

To account for typos in responses, we used the stringdist package (Van der Loo, 2014) in R with the Damerau-Levenshtein distance method. This allowed us to measure the similarity between a participant’s response and the correct answer by computing string distances. Responses with a distance of less than 2 were classified as correct.

To quantify the relative evidence in favour or against our hypotheses of interest, we computed BFs for nested models. If one model includes a parameter reflecting an effect of interest (M1), and a competing model omits this parameter, the BF can be used to quantify the relative evidence in favour or against this effect given the observed data. For instance, a BF₁₀ = 10 for a model including an effect of interest against a model omitting this effect of interest, would mean that the data are 10 times more likely under the model including the effect, compared to a model omitting the effect. Conversely, we can calculate the evidence against the presence of an effect, by computing the reciprocal of BF₁₀. Hereby, BF₀₁ =1/ BF₁₀ would indicate that the data are ten times more likely under the model omitting the effect over the model including the effect. Here, we used the Savage-Dickey density ratio method as an approximation of a nested model comparison for estimating the BF (Wagenmakers et al., 2010). The Savage-Dickey density ratio is calculated as the ratio between the prior and the posterior density of a parameter of interest (i.e., the effect of interest) at a theoretically interesting value. Here, this ratio is calculated for each parameter of interest at the value of 0, as this allows to quantify the evidence that the estimated parameter is different from 0, thereby indicating the presence or absence of an effect.

We implemented Bayesian hierarchical logistic regression models in brms (Bürkner, 2017) to estimate the binary accuracy of strict recall (correct/incorrect). We assumed a Binomial distribution predicted by the model through a logit link function. For all the analyses across the experiments, we included the maximal random-effects structure including random intercepts, as well as random participant effects for all the fixed-effects and their interactions in the model. Cauchy priors with a location parameter of 0 were assigned to all effect parameters. To assess the robustness of our results, we varied the scale of the Cauchy prior (.25, .5, .75, & 1) and re-estimated the BFs five times to guarantee results’ stability. From this, we report the median of the BFs along with the minimum and maximum values.

To test whether there were credible differences between the LTM and new words within both LTM conditions (word pairs and singletons), we fit a model where condition [word pairs, singletons and new] and word type [LTM, new] were included as fixed-effects and then computed the conditional effects for the pairwise comparisons of interest (LTM vs. new words within both the word pairs and singletons conditions). This allowed us to test whether there was credible evidence in favor of an immediate performance benefit for LTM words over new words within trials and whether this was due to a local advantage for episodic representations (benefits only for LTM words that matched pre-learnt word pairs) or to item activation of the words independent of their pre-learnt binding (benefits for both LTM words in the word pairs and singletons conditions).

To test whether there were differences for new words between conditions at each serial position of interest (i.e. proactive and/or retroactive benefits of LTM on WM), we fit three independent models using serial positions [1,2; 2,3; 3,4; 4,5; 5,6]; and condition [word pairs, singletons and new] as fixed-effects. We then computed the conditional effects for the pairwise comparisons of interest that will be described next.

The first model aimed to compare performance for LTM words of the word pairs and singletons conditions to performance of words of the new condition at the same serial positions in which the LTM words were introduced. For instance, immediate memory performance for remembering the LTM words presented at serial positions 1 and 2 to words presented at matching serial positions 1 and 2 in the new condition. This allowed us to test whether an immediate performance benefit for LTM words over new words between conditions was due to a global advantage for episodic representations (an advantage for LTM words that matched pre-learnt word pairs over words from the new condition at the same serial positions) or to item activation of the words independent of their pre-learnt binding (an advantage for both LTM words in the word pairs and singletons conditions over words from the new condition at the same serial positions).

The second and third model aimed to test whether there was a reduction in WM load by including LTM words in WM trials, depending on where the LTM words were introduced within the list. Specifically, the second model tested whether there was a proactive benefit for words introduced after the presentation of LTM words. For example, if LTM words were presented at positions 1 and 2, we analyzed performance differences of words at positions 3, 4, 5, and 6 across all conditions, instead of comparing performance differences to the aggregated performance of all words of the new condition. This allowed for an unbiased comparison focusing only on the serial positions of interest between conditions. The third model aimed to test whether there was a retroactive benefit for words introduced before the presentation of LTM words. Similarly to the second model, we analyzed performance differences across conditions at the serial positions of interest. Hence, if LTM words were presented at serial positions 5 and 6, we analyzed performance differences of words at positions 1, 2, 3 and 4 across all conditions.

Results

Figure 2 shows the proportion of correct responses as a function of (A) condition and (B) for LTM words and new words separately, within the word pairs and singletons conditions. As can be seen in Figure 2B, and supported by the analysis, we observed that immediate memory for LTM versus new information within trials was superior for LTM words that match a pre-learnt word pair in the word pairs condition (BF₁₀ = 1.18 × 10⁵ [1.92 × 10⁴ – 1.15 × 10¹¹]). However, in the singletons condition, when LTM information consisted of two words from different episodes, immediate memory for LTM information was not better compared to new words of the same trials (BF₁₀ = 0.12 [0.07–0.28]).

Figure 2

This means that LTM information included in WM trials benefited immediate memory performance only when the two successive words matched an entire episode stored in LTM. In other words, LTM does not benefit serial recall performance through item activation of each singleton, but rather via the stored binding of the LTM episode.

Next, we examined performance between conditions for LTM, as well as for new information, in order to determine global effects of LTM information, as proactive and retroactive effects depending on the position in which LTM information was presented within the list. Figure 3A shows the serial positions curves for all conditions based on the position of the LTM words within the lists. Figure 3B depicts the comparisons of interest: LTM information of words pairs and singletons conditions compared to words matching the same serial positions of the new condition, and performance of new words following (proactive effect) or preceding (retroactive effect) LTM information (or the matching serial positions of the new condition) within the list.

Figure 3

Supported by the statistical results presented in Table 1, we found a consistent pattern of superior performance for LTM words (in the word pairs condition) at any position within the list compared to words of the new condition at the same serial positions, except at positions 4 and 5. When comparing the word pairs to the singletons condition, there was anecdotal evidence of superior performance for LTM words in the word pairs condition when introduced at the beginning of the list (at positions 1 and 2, 2 and 3) and credible differences at the very end of the list.

The comparison of two words of different episodes (singletons condition) to the new condition only yielded an advantage for the former, when the words were presented at the very beginning of the list (positions 1 and 2).

Regardless of where the LTM words were introduced within the list, whether matching an episode or as singletons, no proactive or retroactive benefit was found for the new words included in the list, meaning that immediate memory performance for new words across conditions was similar irrespective of whether the LTM words were presented before or after them. This suggests that including LTM words at any position of the lists did not free-up WM capacity –meaning it did not result in any advantage or disadvantage for the rest of the words included in the lists.

Discussion

The goal of Experiment 1 was to investigate the flexibility of WM in utilizing LTM representations during serial recall. We aimed to determine whether WM can draw on entire LTM episodes or benefit from the independent activation of each element in the LTM representation when encountered individually. Additionally, we examined whether serial position influences any advantage of LTM words, either through redintegration or reducing WM load.

Our results revealed two main findings: First, we observed superior immediate memory performance for LTM information over new information within and between conditions, consistently only for words belonging to LTM episodes (i.e., pre-learnt word-pairs). Second, incorporating LTM words did not free-up WM capacity, as performance did not improve for novel words of the same lists, regardless of the serial positions of the LTM words.

The advantage of LTM words over words of the new condition was evident when two words from the same LTM episode were sequentially presented compared to introducing two words from different episodes. This speaks against an item activation account and is against our prediction that activating items whether from the same or different episodes would lead to higher performance via redintegration. Instead, it suggests that episodic LTM can benefit WM, when participants can retrieve intact episodic information for serial recall.² Although participants seemed to benefit from matching episodes at any position, there was no benefit for subsequent new information, even when presented at the beginning of the list. This contrasts with previous research showing that semantic chunks introduced at the list‘s onset can free WM capacity (Portrat et al., 2016; Thalmann et al., 2019), but is consistent with previous findings that WM capacity is not freed-up during encoding when two-item chunks are included (Norris et al., 2020).

Overall, the evidence from Experiment 1 is inconsistent with both, the idea that pre-learnt information is chunked or offloaded during encoding, as we did not observe any benefit on new information within mixed lists including LTM information; and the idea that it benefits via redintegration at recall, as our results show that WM is only benefitting from the retrieval of entire episodes but not from item activation.

One possibility to resolve these inconsistencies is to assume that episodic LTM contributions arise from redintegration at recall only for words stored in LTM as full episodes rather than as individual words, thereby facilitating the redintegration effect for intact pairs but not singletons. A trial, which contains an intact pre-learnt word pair could make it more likely that at least some information about this pair is still available at test, thereby increasing the chance for its successful redintegration. Our findings would be consistent with this account, where the advantage is specific to pre-learnt episodes, but there is a lack of a benefit on novel words within the lists (Norris et al., 2020).

Another possibility could be that the current task made it more difficult to properly engage in freeing-up capacity strategies (i.e. chunking or offloading). The current paradigm requires that participants remember each word bound to their positional context. This may limit the usefulness of LTM item-item associations, as recalling each word’s position remains necessary to successfully completing the task. Thus, WM cannot exclusively rely on differently structured LTM traces, limiting the utility of freeing-up capacity strategies at encoding.

A critical difference between this and previous studies in which episodic knowledge improved WM performance is the level of facilitation in recognizing LTM information. Previous studies used tasks that inherently facilitated the recognition of pre-learnt episodes during encoding by utilizing the same item-item structure (Bartsch & Shepherdson, 2022, 2023; Chen & Cowan, 2005, 2009) or clearly separating pre-learnt information from new information (Thalmann et al., 2019). In our study, words that matched prior knowledge were less salient. By presenting them centrally like new words, we potentially reduced the effectiveness of chunking or offloading strategies that are ought to occur at encoding. Therefore, an alternative explanation for the lack of freeing-up WM capacity benefits in our results is that the different structures of LTM and WM representations made it harder to immediately recognize matching pre-learnt episodes. As a result, participants may have had to initially encode both words independently, limiting their ability to engage in encoding strategies that could otherwise reduce WM load.

Indeed, recent evidence highlights the role of awareness at encoding in WM for retrieving episodic memory. Specifically, a recent study on Hebb repetition learning (Musfeld et al., 2023), demonstrated that participants only benefited from repetitions once they had recognized a repeating episode during re-encoding. Thus, recognizing previously encountered information during WM encoding is critical for learning and retrieving LTM information.

In summary, the different representational structures used in WM and LTM might have prevented participants from recognizing and subsequently chunking or offloading redundant information during encoding. Instead, they likely encoded the information as item-context bindings, preventing the freeing of WM capacity and resulting in no advantage for new information. We test this possibility in Experiment 2.

Method

Data analysis

The data were analysed equivalently to Experiment 1, but this time the fixed-effect of serial position included an additional level of serial positions 6 and 7 [1,2; 2,3; 3,4; 4,5; 5,6; 6,7] and the fixed-effect of word type included the highlighted red words and neutral black words [red, black]

Results

Figure 4 shows overall performance across (A) conditions and (B) distinguishing performance depending on the color in which the words were introduced.³ In contrast to Experiment 1, this time the advantage of LTM information over new information was found not only within the word pairs condition (BF₁₀ = 1.82 × 10⁵ [6.08 × 10⁴ – 4.79 × 10⁶]), but also within the singletons condition (BF₁₀ = 3.21 [1.60 – 5.72]) (see Figure 4B). The latter reflected an item activation boost for LTM words (yet which were not associated within the same episode). Importantly, there was no effect of saliency when comparing red versus black words within the new condition (BF₁₀ = 0.21 [0.11 – 0.50]).

Figure 4

Based on the serial position curves from Figure 5A, we now compare performance between conditions of LTM versus new information, and to test proactive and retroactive effects on new information based on the position of LTM words within the list (see Figure 5B). First, and supported by the statistical results presented in Table 2, immediate memory for words from pre-learnt word pairs were remembered consistently and credibly better at all serial positions compared to words of the new condition presented at the same serial positions. Moreover, compared to the singletons condition, immediate memory for LTM words from the word pairs conditions was credibly superior at the beginning and at the very end of the list (i.e., positions 1 and 2, 2 and 3, and 6 and 7). In contrast, there were no credible differences between singletons and words of the new condition at any serial positions within the list.

Figure 5

Regarding proactive and retroactive effects, our findings indicate once again that the observed improvement in performance for LTM words did not translate into an advantage for the performance of novel words presented either prior or subsequently to the LTM words compared to the new condition, irrespective of the position in which the latter were introduced within the list.

Discussion

The main goal of Experiment 2 was to investigate the possibility that participants were unable to recognize the matching episode at encoding due to the lack of a 1-to-1 correspondence in the structure of the WM representations – subsequently preventing any effects of freeing WM capacity. To achieve this, we decided to guide and facilitate awareness of LTM information, by highlighting LTM words in red and informed participants that this could be indicative of words they had pre-learnt during the LTM learning phase.

First, our results showed that LTM words of the word pairs condition had a better overall performance compared to the singletons and new condition, which replicates the observed pattern in Experiment 1.

Second, manipulating the saliency of the information by making participants aware that the color was informative, not only resulted in consistently superior performance of LTM words compared to new words within the word pairs condition, but also extended to LTM singletons. However, when comparing performance of LTM words to words of the new condition at all the same serial positions, our results revealed superior performance of LTM words in the word pairs condition regardless of their serial positions within the list. This was not observed when comparing LTM singletons to words of the new condition at any serial position.

While the saliency manipulation boosted performance through item activation of the singleton LTM words compared to new words in the same list (local effect), immediate performance for sequentially presented singletons did not exhibit an advantage when compared to new words of the new condition at any serial position (global effect). This indicates once again that superiority of performance for LTM words, both locally and globally, is more robust only when both LTM words form part of a pre-learnt word pair, consistent with Experiment 1.

Third, although memory for the LTM words themselves was improved, our results again provided no evidence of freeing WM capacity. Instead, they are consistent with a redintegration account, when assuming that this process is more likely to occur for intact episodes rather than single words. The absence of any advantage for new information in lists containing LTM words suggests that participants did not reduce WM load when LTM information was included, not even in the form of pre-learnt word pairs. This lack of freeing-up capacity indicates that information was probably not successfully recoded or offloaded during encoding, not even when the information was presented at the very beginning of the list. This stands once again in contrast to previous research using semantic LTM representations (Portrat et al., 2016; Thalmann et al., 2019) or larger pre-learnt lists of at least three elements (Norris et al., 2020).

We argue that the creation and maintenance of positional bindings in intermixed lists for both, LTM and new words, is crucial for successfully completing the serial recall task. This is particularly relevant in the context of LTM representations lacking positional information, especially for paired associates with symmetrical properties (Kahana & Caplan, 2002), where no ordinal or positional information is inherent to the representations. As a consequence, the implementation of any chunking strategy during encoding or purely relying on the LTM representations (i.e. offloading) is not sufficient to successfully solve the serial recall task.

Beyond the demand of encoding positional bindings, an additional challenge may arise when pre-learnt word pairs must be encoded separately without structural support at WM encoding. Such support can be observed when the word pairs are presented simultaneously (Bartsch & Shepherdson, 2022, 2023; Chen & Cowan, 2005, 2009), or when a grouping structure is used. Previous studies have achieved this by organizing the information spatially, for example, by placing the LTM information in separate rows from new lists (Musfeld et al., 2024; Thalmann et al., 2019).

Therefore, the utilization of a clearly defined structure that can be employed during both encoding and retrieval may facilitate not only the recognition of the LTM representation, but also might serve as a template for recalling information that already exists, allowing for a more efficient reallocation of resources and freeing of capacity in WM. To test this hypothesis, we ran Experiment 3.

Method

Results

Figure 7 depicts the proportion of correct responses as a function of (A) condition and (B) performance for LTM and new words independently within both LTM conditions (intact vs. disrupted). As shown in Figure 7B and supported by the analysis, performance of LTM words was credibly higher compared to new words within trials irrespective of the encoding structure in which the LTM words were introduced in the WM templates (intact condition: BF₁₀ = 2.06 × 10⁶ [1.72 × 10⁵ – 1.47 × 10⁷] ; disrupted condition: BF₁₀ = 1.43 × 10⁵ [3.80 × 10⁴ – 7.36 × 10⁵]).

Figure 7

We next focused on performance between conditions for LTM word pairs and words from the new condition based on the position of word pairs within the lists (see Figure 8). As shown in Figure 8B and supported by the statistical results presented in Table 3, we consistently found superior immediate memory performance for LTM word pairs compared to words in the new condition at all serial positions (beginning, middle and end of list), and regardless of the encoding structure (intact or disrupted).

Figure 8

While descriptively it appears that the intact encoding structure enhanced performance of word pairs compared to those encoded in a disrupted structure, the only credible difference in which memory performance for word pairs of the intact condition was higher compared to those of the disrupted condition was when they were presented at serial positions 3 and 4. This suggests that the benefits of presenting a 1-to-1 matching structural representation to that in which the episode was learnt are most pronounced at the beginning of the list. However, presenting information at the very beginning (i.e. the edge) of the list leads to equivalent performance independent of the visuo-spatial arrangement in which the LTM pairs were introduced.

Notably, for the first time in these set of experiments, we observe a benefit for new words when LTM pairs were introduced within a list – indicating freeing of capacity. When the LTM information was presented at the very beginning of the list, we observed a proactive memory benefit for new words (compared to the new words of the new condition at the same serial positions), irrespective of the encoding structure (intact vs. disrupted; see Table 3). For the intact condition, this proactive benefit was also present when the LTM pair was presented at positions 3 and 4, but no longer for the disrupted condition. When the LTM pair was presented towards the end of the list, no proactive effect was present regardless of the format. None of the conditions produced a retroactive benefit (see Table 3).

Discussion

Our main objective in Experiment 3 was to test whether incorporating LTM word pairs in a serial recall task, using a visuo-spatial configuration in WM that matched the learning phase, would not just ease recognizability and recall of the LTM word pairs, but also reduce WM load.

To achieve this, we used visuo-spatial templates wherein the LTM word pairs could be presented adjacent to each other forming a group (intact condition) – while still being sequentially presented. This mirrored the structure used during the learning phase. We contrasted this with a disrupted structure, where the two words forming the word pair were visuo-spatially separated.

Our results revealed that performance of LTM word pairs was consistently higher compared to new words within the same lists, regardless of whether the encoding structure in the template matched the pre-learnt LTM representations (intact condition) or whether it was disrupted. Moreover, when compared to words of the new condition, LTM word pairs showed superior performance across all the matching serial positions (beginning, middle, and end of the list).

Furthermore, the utilization of visuo-spatial templates during both encoding and recall in WM not only resulted in benefits for the LTM words themselves but also contributed to reducing the load on WM, particularly when LTM pairs were introduced at the very beginning of the list in both intact and disrupted encoding structures. This advantage only extended further to conditions in which the LTM pair was presented later in the list when the encoding structure was intact, but not when it was disrupted.

Compared to Experiment 1 and 2, this experiment created a context in which information was visually organized at encoding. This manipulation not only facilitated the recognition of LTM information and enhanced its immediate recall, but also contributed to freeing-up WM capacity for subsequent new words, although only when WM was not yet loaded. This is consistent with previous findings using semantic LTM representations (Mızrak & Oberauer, 2021; Portrat et al., 2016; Thalmann et al., 2019).

The fact that the proactive effects were observed for both intact and disrupted conditions – in case the word pairs were presented at the beginning of the list – suggests that an identical matching encoding structure may not play a major role at this edge position within the list. This is because the information presented at the beginning of the list is inherently distinct and salient (Burgess & Hitch, 2005; Henson, 1999). However, the observed advantage of the intact structure over the disrupted one, at later positions (3 and 4) speaks to the benefits of encountering a matching visually structured episodic representation. This format most likely facilitated recognition of prior learnt representations and enabled a more effective engagement in chunking or offloading strategies to release resources for upcoming information, as evidenced in our results.

The revealed proactive benefits shown here, are in line with the assumption that LTM information would aid in the efficient processing of new stimuli once it has been chunked in LTM, by which redundant information is removed from WM and does not interfere with upcoming information (Mızrak & Oberauer, 2022; Norris et al., 2020; Norris & Kalm, 2021; Thalmann et al., 2019). This could entail a process such as offloading the second word associated with the first one of the word pair, probably after both elements have been encoded. We further argue that fully offloading the word pair is unlikely, because of the reason mentioned above: the task requires the creation of positional bindings to solve the task, as at least one element needs to be bound to the specific context (the position within the list).

As suggested by Thalmann, et al. (2019), when the LTM pair appears later within the list, previously encoded representations may be adversely affected by the encoding of subsequent LTM elements before they can undergo the chunking process. Consequently, as the list lengthens and more information is retained in WM, the potential harm increases, thereby impeding the ability of LTM information to mitigate this damage.

These results indicate overall that the way we structure and encode the information in WM is an important factor that regulates the effectiveness of LTM contributions to WM.

Benefits of episodic LTM to WM

The consistent finding across these and previous experiments incorporating prior knowledge in the form of episodic LTM representations in WM tasks, is the emergence of a clear benefit for LTM representations over novel information within the same lists (Bartsch & Shepherdson, 2022, 2023; Norris et al., 2020; Thalmann et al., 2019). In our case, this mainly arose in case the sequentially presented LTM words consisted of a pre-learnt episodic association in the context of serial recall.

However, despite the consistent advantage for recalling pre-learnt information over newly presented information, we did not consistently observe freeing of WM capacity for new words. Why is a capacity-freeing effect not consistently observed across different WM tasks?

Considering our foundational understanding of WM as a limited capacity system primarily focused on holding relevant information to fulfill current demands (Cowan, 2008; Oberauer, 2009), the goal of the WM task itself becomes crucial in regulating the system’s flexibility in relying on prior information. Granting access to LTM information through a flexible gate, as previously suggested (Bartsch, Frischkorn, et al., 2024; Bartsch & Shepherdson, 2022; Mızrak & Oberauer, 2022; Oberauer et al., 2017), therefore must hold benefits for solving the task at hand – over and above the costs that might arise from sources of interference. Therefore, when previously learnt information contains the necessary bindings to accomplish the WM task at hand and can be reliably retrieved during recall, drawing on LTM should be advantageous. However, when such bindings are not inherently included in the LTM representations, WM must create the necessary temporal bindings of information to the context from scratch (Oberauer, 2019).

In the present study, the pre-learnt item-item associations, represented as word pairs, would have required additional bindings to the context to successfully complete the serial recall task. Therefore, the LTM words may not have undergone a chunking or offloading process, preventing freeing of WM capacity. Still, even in this situation, a benefit for the LTM information itself could nevertheless arise. For instance, the existing LTM representations could help to reconstruct the WM trace at recall, in case the trace was degraded (Botvinick, 2005; Hulme et al., 1997; Norris et al., 2020; Thorn et al., 2005) compared to weaker representations of newly encountered information that have no support of pre-learnt episodes.

Indeed, when inspecting previous tasks in which a capacity-freeing effect of LTM was observed, we find that creating bindings from scratch for information included in WM was not inherently necessary. This is the case for tasks which did not require the formation of new associations to the context, such as recalling the paired word in a cued-AFC task (Bartsch & Shepherdson, 2022, 2023); or in cases in which the task required the creation of fewer bindings (Chen & Cowan, 2005, 2009). In the latter studies, when pre-learnt word pairs were presented simultaneously in pure LTM pair lists, recognizing the pair might have allowed to only encode and bind the first word to its positional context. For the second word, no additional binding was required, as it could be recalled based on its association to the first word during recall (Chen & Cowan, 2005, 2009). Lastly, other previous tasks used an encoding format which allowed for chunking strategies at recall, such as presenting chunks as independent lists on separate rows of the screen (Thalmann et al., 2019). Here, retrieval of the full list only required encoding of the first word (always associated with serial position 1) of the LTM list. Hence, the need to create new bindings for all the elements within WM may have been circumvented.

Encoding structure matters

In the present study, a reduction in memory load emerged when the serial recall task was supported by the utilization of visuo-spatial templates to organize the to-be-encoded information. Recent findings have highlighted the importance of encoding structures for learning repeated lists: they facilitate the encoding of information into an integrated representation (Musfeld et al., 2024) and changes in the structure at encoding have been shown to have detrimental effects in learning new associations (Bower & Winzenz, 1969; Musfeld et al., 2024).

We argue that the use of such structures facilitates not only awareness of matching pre-learnt episodes – a necessary, though not sufficient step, as demonstrated in Experiment 2 – but also a more efficient way to rely on visuo-spatial templates for better reallocation of resources, therefore a reduction in WM load when encountering episodic LTM representations. This, importantly, depended on where the information was presented within the list, and was more pronounced when the visuo-spatial template matched the representation in which the LTM word pairs were acquired (intact condition). This pattern aligns with Thalmann et al.’s (2019) findings, where proactive rather than retroactive effects are observed using semantic chunks: the benefits of memory load reduction appear when WM capacity has not been exceeded. This is because encoding the elements of the word pair individually after new words have already been encoded could damage these representations, as earlier presented items may be more degraded and susceptible to interference. However, engaging in a chunking strategy earlier in the list for LTM information allows for a reduction in WM load, as long as the information can be chunked or offloaded – for serial recall, presumably partially offloaded. Having the visuo-spatial template as additional support allowed participants to not only rely on item-position bindings but also to derive extra benefits from the spatial context. In this case, the easier an episode is recognized, the easier participants can chunk the elements, and the less resources are required. Hence, more resources will be freed-up and are available to items following from pre-learnt LTM pairs. This is consistent with the suggestion by Norris et al. (2020) that the efficiency of engaging in chunking strategies is dependent on the costs of recoding or offloading LTM information during encoding. However, our findings show that these costs are not only dependent on the size of the chunk as shown by Norris et al. (2020), but also on the ease of recognizing and matching an existing LTM representation to the current encoding context.