Forming memories of past events or learned experiences is one of the cognitive abilities that humans deeply rely on. Ebbinghaus (1885) showed that forgetting occurs very quickly after new learning without reviewing or reinforcing the learned contents. What causes forgetting is a question that has attracted the attention of many cognitive psychologists. Before 1960, interference was supposed to be the main mechanism for forgetting (Nairne, 2003). The notion that mnemonic information is spontaneously lost as time passes by had fallen out of favor (at least for the long-term retention) primarily for two reasons: a) memory sometimes improves with time, and b) forgetting depends vitally on the nature of interfering material. As McGeoch (1932) noted, the process of forgetting may be similar to that of rusting of an iron bar in open air: rust accumulates with time, but it is the process of oxidizing, not time per se, that is responsible for the changes in the iron.

Later studies suggested that forgetting of short-term retention is not entirely due to interference. The critical evidence came from the single-trial immediate serial recall tasks, with the Brown-Peterson technique (Brown, 1958, Peterson and Peterson, 1959). In the Brown-Peterson procedure, in each trial, a single consonant trigram, such as TSF, is shown and recalled, with a retention interval in between the presentation and recall phase. The participant was asked to count backwards aloud (by threes) throughout the interval. The finding was that the consonant trigram was essentially forgotten after a retention interval of 18 s. Because not much interference was expected to occur between highly dissimilar materials (such as between consonant trigram and numbers), the observation of significant forgetting after only 18 s interval demonstrated a fast forgetting in short-term memory that cannot easily be explained by the inter-item interference, especially given that the memory load was well below span, i.e., only one consonant trigram.

The idea that both decay and interference are involved in short-term memory had largely dominated the field for several decades since the 1970s. According to this view, successful retention depends not only on the ability to keep memory items in an active and recallable state, but also on the ability to remove interference from the nontarget information in memory. It has been suggested that attention is used to inhibit or reject interfering material as well as to keep target items active in memory (Dempster, 1992, Kane and Engle, 2000).

Recent models (especially the computational models) of working memory1 started to emphasize the dynamic role that attention plays in the working memory retention. Specifically, in the past two decades, there has been a very heated debate on whether the dynamic role of attention is to resist decay (by refreshing memory representations) or to remove interference (Barrouillet and Camos, 2015, Barrouillet et al., 2004, Oberauer et al., 2012). A typical decay-based theory, e.g., the TBRS (time-based resource sharing) model, posits that representations in working memory decay with time, unless being refreshed by the focus of attention (Barrouillet et al., 2004). A typical interference-based theory, e.g., the SOB-CS (serial-order-in-a-box, complex span) model, claims that interference from non-memory items causes the forgetting, but the interference can be eliminated by an attentional process (Oberauer et al., 2012).

Empirical evidence supporting both decay-based and interference-based theories came from complex span tasks that interleave the to-be-recalled material (e.g., letters) with processing tasks (e.g., counting) and assess the effect of the cognitive load of the processing task on the memory performance (Fig. 1). Cognitive load was defined as the ratio of the minimum time required to fulfill the processing task to the time interval between two consecutive memory items. The common finding is that working memory performance (e.g., the memory span) linearly decreases with the cognitive load of the processing task.

Both theories claim to explain the “cognitive load effect”. According to TBRS model, the higher the load, the less resources could be used for attentional refreshing, causing a reduction in the number of memory items that can be maintained by the attentional refreshing mechanism (Barrouillet, Portrat, & Camos, 2011). By contrast, SOB-CS model suggests that the higher the load, the less time is available to remove the interference, leading to a reduction in the number of memory items that can be recalled (Oberauer et al., 2012). The debate between TBRS and SOB-CS is still unresolved, mostly because previous studies seldom directly manipulated decay or interference. They manipulated mostly the cognitive load by varying the proportion of time attention is occupied during the inter-memory-item intervals. However, effect of time only offers indirect evidence, because time can be used for refreshing memory presentations as well as for removing interference.

It is important to note that the decay mechanism and the interference mechanism are not mutually exclusive. Instead, they may both exist. As shown in Fig. 2, attention may be able to resist decay via the attentional refreshing mechanism and may also be able to eliminate interference via the interference removal mechanism. Since the time available for attention can be controlled by manipulating the cognitive load, the effect of cognitive load could be due to the refreshing mechanism, the removal mechanism, or both. Therefore, it is impossible to verify the refreshing mechanism and the removal mechanism at the same time by manipulating the cognitive load. The two mechanisms need to be verified one by one. In order to testify the refreshing mechanism, we need to independently manipulate the time available for attention (cognitive load) and the strength of decay, and see whether there is an interaction between the effects of these two variables on recall performance. This is because if the effects of these two variables on recall performance are independent, it suggests that the time available for attention (cognitive load) cannot affect the decay process via the assumed refreshing mechanism. Similarly, to testify the removal mechanism, one also has to independently manipulate the time available for attention (or cognitive load) and the amount of interference, and see whether there is an interaction between the effects of these two variables on recall performance. If the effects of the two variables on recall performance are independent, it means the time available for attention cannot affect the interference via the assumed removal mechanism.

Because it is difficult to independently manipulate the time available for attention (or cognitive load) and the strength of decay, in the present study, we mainly examined the removal mechanism by independently manipulating the time available for attention (or cognitive load) and the amount of interference. In the complex span task, there are two types of interference that may be produced by the concurrent processing task: the storage interference and processing interference. The storage interference is produced by the storage competition, i.e., items of the memory task need to compete with items of the processing task that are encoded into working memory for the storage resources. By contrast, the processing interference is due to the processing competitions, i.e., the central executive in working memory not only is responsible for the memory task (e.g., encoding, consolidating and maintaining the memory items), but also is involved in the processing task, so that the memory processing needs to compete with the concurrent non-memory processing. The processing interference and attentional resources cannot be manipulated independently, because the processing interference itself stems from competition for attentional resources. Thus, in the present study, we tested the validity of the removal mechanism by independently manipulating cognitive load and amount of storage interference.

For simplicity, we call the assumption that “storage interference from concurrent processing tasks can be removed by attention” the removal hypothesis, and we call the assumptions of the TBRS model the attentional refreshing hypothesis. In fact, previous studies have already questioned the removal hypothesis, but those studies were not sufficient to challenge it due to their own drawbacks. Here, we first give a brief review of those studies.

Some opponents of the removal hypothesis have modified the complex span tasks, trying to differentiate the predictions between the removal hypothesis and the account of attentional refreshing in order to better testify the removal hypothesis. For example, Barrouillet, Portrat, Vergauwe, Diependaele, and Camos (2011) attempted to hold constant the cognitive load and the total amount of interference from the processing task, while varying the time interval between consecutive memory items. If the removal hypothesis holds, the longer interval between consecutive memory items, the more interference should be removed, and thus the better the recall performance of the memory task, because the total amount of interference was held constant. In contrast, the account of attentional refreshing predicts that the mechanism of attentional refreshing is influenced only by cognitive load, so that the manipulation of interval between the consecutive memory items should not affect recall performance since the cognitive load was held constant. Barrouillet et al. found that this interval manipulation did not affect recall performance and argued that their results supported the attentional refreshing account but not the removal hypothesis.

However, Barrouillet et al.'s experimental control may not have been as successful as expected. In their task, participants were required to remember a series of letters. In between the consecutive letters, participants performed several position judgement tasks, in which they judged whether a square was presented in the upper or lower part of the screen. The number of position judgment tasks within the interval between each pair of consecutive memory items was manipulated. The cognitive load of the position judgment task was held constant so that the inter-memory-item interval increased with the number of processing tasks. Because they presented squares that appeared exactly the same in all the position judgment tasks, they assumed the amount of interference was constant. However, this assumption was questionable. As pointed out by Oberauer and Lewandowsky (2014), after interference from a processing task is removed, the next processing task will reintroduce the same interference that needs to be removed again. Therefore, the results of Barrouillet et al. may in fact be consistent with the predictions of the removal hypothesis, because the increased interval was mostly used to remove the reintroduced interference, rendering the manipulation of interval ineffective on the recall performance.

Opponents of the removal hypothesis have also proposed another way to test the removal hypothesis, i.e., by directly examining the recall performance of the interfering items (to determine whether the interfering items were actually removed). According to the removal hypothesis, the longer interval between consecutive memory items, the more completely the interfering items could be removed, and therefore, the worse the recall performance of the interfering items. Based on this reasoning, some studies have directly examined the recall performance of the interfering items in complex span tasks (Dagry and Barrouillet, 2017, Dagry et al., 2017). For example, Dagry et al. (2017) asked participants to complete a complex span task that was composed of a number of trials. A period after the completion of all trials, the participants were required to freely recall all the memorized and non-memorized items they saw in the complex span task. It was shown that the manipulation of inter-item interval between consecutive memory items did not affect the number of non-memory items recalled. If there were an interference removal mechanism, longer interval should allow more non-memory distractors to be removed and, thus, lead to a reduction in the number of non-memory items eventually being recalled. Therefore, Dagry et al. concluded that their results were not consistent with the predictions of the removal hypothesis. However, as Lewis-Peacock, Kessler, and Oberauer (2018) pointed out, it is possible that the removal hypothesis only predicts recall performance on working memory tasks but not on long-term memory tasks. Since the study of Dagry et al. bears on the long-term memory performance, it cannot be used to challenge the removal hypothesis.

Perhaps Barrouillet and colleagues had foreseen the above counterargument, and so, Dagry and Barrouillet (2017) modified the task of Dagry et al. (2017) by replacing the delayed recall with the immediate recall (i.e., requiring participants to recall all memory items with distractors immediately after each trial). It was shown that a longer interval between the consecutive memory items increased rather than reduced the number of recalled non-memory items. Clearly, this result is also inconsistent with the prediction of the removal hypothesis. However, Lewis-Peacock et al. (2018) pointed out that the interference removal mechanism may only remove the bindings between the memory items and their context (e.g., retrieval cues), rather than the memory items per se (see Oberauer et al., 2012). Because Dagry and Barrouillet used the free recall task which cannot determine whether the bindings between memory items and their context (e.g., retrieval cures) were removed, their results still cannot provide strong challenge to the removal hypothesis.

In summary, there is still a debate about the validity of the removal hypothesis. An important reason for the difficulty in ending this debate is that, in the relevant studies, the total amount of interference introduced by the processing task was often covariant with the amount of the attentional resources available to remove interference, i.e., the interference has never been successfully manipulated independently of the attentional resources. If the amount of interference can be controlled independently of attentional resources, it would allow us to examine whether the interference removal mechanism exists (see Fig. 2).

In the present study, we aimed to verify the removal hypothesis by independently manipulating the storage interference and attentional resources (by altering cognitive load). To accomplish this, we borrowed ideas from Li and Li (2021) study. Li and Li (2021) found little storage interference between stimuli from different domains (e.g., color vs. letter) even under high cognitive load. They also found significant storage interference between a processing task and a memory task if stimuli involved in the two tasks share features in the same domain. For example, in Li & Li’s Experiment 4, participants were asked to detect specific target letters in an RSVP (rapid serial visual presentation) letter stream (all letters were white) while memorizing five colors. They found that the letter detection task caused little storage interference with the color memory task. Moreover, in Li & Li’s Experiment 5, participants did a single-target tracking task while memorizing the color and position of three color-dots. In the retrieval phase, either the color or the location of one of the dots was recalled. The results showed the tracking task produced significant storage interference to the spatial memory task but no storage interference to the color memory task. These results suggest that a concurrent processing task may produce storage interference to a memory task only if the two tasks share features in the same domain.

Therefore, in the present study, we used a similar design as used in Li & Li’s study, requiring participants to concurrently perform a RSVP letter detection task and a color memory task. The key modification in the present study was that the letters in the RSVP stream were either black or colorful. We reasoned based on Li & Li’s findings that when the letters were black, the letter detection task may produce little storage interference to the color memory task, whereas when the letters were colorful, the letter detection task may produce strong storage interference to the color memory task. That is, in the present study, the storage interference was controlled by the degree of color involved in the RSVP letter detection task (i.e., colorful letters correspond to high interference; black letters correspond to low interference). The attentional resources (cognitive load) were manipulated by the number of letters per unit of time in the RSVP letter stream (i.e., multiple letters in the RSVP stream corresponds to high load, whereas single letter in a stream corresponds to low load). Since Li & Li has shown that the white RSVP letter stream cannot produce storage interference to color memory, the experimental design of the present study allows for the independent manipulation of attentional resources and storage interference, and thus it has the potential to determine the validity of the removal hypothesis.