Introduction

Working memory (WM) enables the temporary storage and manipulation of information no longer in the sensorium. This system has a limited capacity (e.g., Luck & Vogel, 2013; Ma et al., 2014), and mechanisms of selective attention are needed to control what information gains access to WM and to prioritize existing WM representations for behavioral output. It is well established that WM performance can be facilitated by an informative cue presented during storage (e.g., Griffin & Nobre, 2003; Landman et al., 2003), and this effect – termed a retrocue benefit – can be used to explore the behavioral and neural consequences of prioritizing information stored in memory (e.g., Ester et al., 2018; Myers et al., 2017; Nouri & Ester, 2020; Souza & Oberauer, 2016; Sprague et al., 2016).

Several studies have demonstrated that human observers can use spatial information to flexibly shift attention between different items stored in WM. For example, Landman et al. (2003) presented human volunteers with sequential location retrocues while they remembered arrays of oriented stimuli, and found a retrocue benefit for the most recently cued orientation (see also Li & Saiki, 2014; Maxcey et al., 2015). Using a similar procedure, Souza et al. (2015); see also Rerko & Oberauer, 2013) found a retrocue benefit for sequentially cued stimuli at multiple locations, provided that each cued stimulus was equally likely to be tested at the end of a trial. Thus, participants can utilize spatial information to sequentially prioritize different representations stored in WM to improve memory performance. Whether participants can use spatial information to simultaneously prioritize WM content is less clear. On the one hand, some studies have found that multiple simultaneously presented location cues confer no benefit on WM performance compared to a no-cue or neutral cue condition (Makovski & Jiang, 2007; Oberauer & Bialkova, 2009). On the other hand, some studies have found that multiple simultaneously presented location cues confer a benefit on WM performance during some conditions (e.g., when the retrospectively cued stimuli appeared in different visual hemifields, or when the retrospectively cued stimuli are adjacent; e.g., Delvenne & Holt, 2012; Heuer & Schubö, 2016; Matsukura & Vecera, 2015), but not during other conditions.

Many studies examining simultaneous spatial retrocue benefits on WM performance have relied on aggregate measures of memory performance such as average change detection accuracy or absolute recall error. These measures have the advantage of simplicity but make it difficult to determine how cues influence memory performance. For example, location retrocues could influence WM performance by (a) increasing recall precision for cued relative to uncued memory stimuli, (b) decreasing the likelihood of forgetting a cued versus an uncued stimulus, (c) decreasing the likelihood of confusing the cued stimuli with uncued stimuli (i.e., a swap error), or (d) some combination of the above. Indeed, several studies have reported that compared to a neutral- or no-cue condition, a single location retrocue improves recall precision, lowers swap rates, and lowers the probability of random guessing (e.g., Gunseli et al., 2015; Makovski & Pertzov, 2015; Pertzov et al., 2013). What happens when participants are retrospectively cued to prioritize stimuli that appeared in different locations? It is possible that these cues confer benefits on some aspects of memory performance (e.g., reducing the likelihood of random guessing) while harming other aspects of memory performance (e.g., increasing the likelihood of swap errors). However, this pattern would be opaque to discretized or aggregate measures of memory performance. To illustrate, consider a hypothetical experiment where participants encode four items into memory and are subsequently probed to recall the identity of one item. During storage participants receive a spatial cue indicating that the to-be-probed item will be drawn from a subset of two items stored in memory (i.e., informative cue trials) or a cue indicating that all four items are equally likely to be probed (i.e., uninformative cue trials). Suppose that the results of this experiment show that during uninformative cue trials participants correctly recall the identity of the probed item with probability 0.7, incorrectly recall the identity of a non-probed item (i.e., a task-irrelevant stimulus) with probability 0.1, and randomly guess with probability 0.2, but during informative cue trials participants recall the identity of the probed item with the same probability of 0.7, but recall the identity of a non-probed item with a greater probability of 0.25, and randomly guess with a lower probability of 0.05. Clearly, these patterns would indicate that participants are processing stored information differently during informative and uninformative cue trials, yet they would yield (nearly) identical average absolute recall error estimates.

We hypothesized that human volunteers can utilize multiple simultaneously presented spatial retrocues, but that doing so induces a trade-off between guessing and swap errors. Specifically, we predicted that retrospectively cueing participants to prioritize two stimuli that appeared in different locations would lead to a reduction in random guessing during informative versus neutral cue trials. We also predicted that retrospectively cueing participants to two prioritized stimuli that appeared in different locations would increase competition or inter-item interference between the prioritized memory representations, increasing the likelihood that the prioritized items are swapped and leading to a greater proportion of non-target reports during informative versus neutral cue trials. We tested these predictions in three experiments. Participants received location retrocues instructing them to prioritize oriented (Experiment 1 or Experiment 2) or colored (Experiment 3) stimuli. Analyses of participants’ average absolute recall errors revealed a significant performance benefit during cue-one relative to cue-two and cue-zero trials, replicating several earlier findings (e.g., Makovski & Jiang, 2007; Oberauer & Bialkova, 2009). Analyses of participants’ recall precision, swap rates, and guess rates revealed that superior recall performance during cue-one relative to cue-two and cue-zero trials was driven primarily by lower guessing rates, consistent with other prior findings (e.g., Murray et al., 2013; Pertzov et al., 2013). However, precision estimates, swap rates, and guess rates were identical during cue-two and cue-zero trials. Thus, we found no evidence to support the hypothesis that cuing participants to prioritize stimuli that appeared in different locations improved memory performance compared to an uninformative cue display or induced a trade-off between random guessing or memory confusions. These results, in turn, lend additional support to the hypothesis that location-based access to the contents of WM is limited to a single stimulus at a time (Makovski & Jiang, 2007; Oberauer & Bialkova, 2009; Souza & Oberauer, 2016).

Methods

Participants

A total of 98 volunteers participated in this study. Forty-five volunteers from the Florida Atlantic University community participated in Experiment 1, 28 volunteers from the University of Nevada, Reno community participated in Experiment 2, and 25 volunteers from the Brock University community participated in Experiment 3. All participants were aged 18–40 years and self-reported normal or corrected-to-normal visual acuity. All experimental procedures were approved by local institutional review boards, and all volunteers gave both written and oral informed consent before enrolling in the study. Data from eight participants in Experiment 1 were discarded due to chance-level task performance (i.e., average absolute recall errors ≥ 85° in the cue-one condition). Thus, the data reported here reflect the remaining 90 participants (37 in Experiment 1, 28 in Experiment 2, and 25 in Experiment 3).

Stimulus displays and testing environment

Participants in each experiment were seated in a dimly lit and quiet room for the duration of testing. Stimuli for Experiments 1 and 2 were generated in MATLAB and rendered using Psychtoolbox 3 software extensions. Stimuli in Experiment 1 were rendered on a 19-in. Dell CRT monitor cycling at 75 Hz; stimuli in Experiment 2 were rendered on a 27-in. LCD monitor cycling at 240 Hz. Participants were seated approximately 65 cm from the display (head position was unconstrained). Stimuli for Experiment 3 were generated in Python and rendered on a 20-in.’ LCD display using PsychoPy software (Peirce et al., 2019). Participants were seated approximately 57 cm from the display (head position was unconstrained).

Experiment 1: Exogenously cued orientation recall

A trial schematic is shown in Fig. 1A. Participants were instructed to maintain fixation on a small dot (subtending 0.25° visual angle from a viewing distance of 60 cm) for the duration of each trial. Each trial began with a sample display containing four “clock-face” stimuli at 45°, 135°, 225°, and 315° polar angle along the perimeter of an imaginary circle (radius 5°) centered at the fixation point. Each stimulus subtended 2.5° (diameter) and contained a bar (1.25° length, 8-pixel stroke width) whose orientation was randomly and independently chosen from a uniform circular distribution on the interval (0°, 359°). The sample display was presented for 500 ms and followed by a 1,000-ms blank delay. A cue display was presented for 100 ms, followed by a 400-ms blank delay and a probe display containing a single clock-face stimulus. The probe stimulus was assigned a random orientation value, and participants were instructed to adjust it to match the sample stimulus it replaced using the left and right arrow keys. Participants entered their final response by pressing the spacebar. Participants were instructed to prioritize accuracy, and no response deadline was imposed.

Fig. 1
figure 1

Design and Results of Experiment 1. A Task schematic showing a cue-two trial. Displays have been enlarged for exposition; see Methods for exact parameters. B-E Average absolute recall error (B), estimated recall precision (C), swap rates (D), and guess rates (E) as a function of cue condition. Error bars depict the 95% confidence interval of the mean

Full size image

Spatial cues presented during WM storage instructed participants to retrospectively prioritize zero, one, two, or all four stimuli. Cues were rendered by flashing the circular outline of the relevant stimuli white for 100 ms (see Fig. 1A). During cue-one trials we randomly cued one of the four stimuli, subject to the constraint that each location was cued equally often within a single block of 60 trials. During cue-two trials we randomly cued two of the four stimuli. We did not explicitly control the spatial relationship between the cued stimuli, i.e., whether they appeared in the same versus different hemifields, but we did investigate possible effects of cue location in post hoc analyses (e.g., Fig. 2). The cue-zero and cue-four conditions were included as neutral baselines. Both conditions yielded equivalent performance, so data from these trials were pooled to create a single neutral cue condition (specifically, we analyzed the cue-zero and cue-four trials separately for each observer and then averaged the data across conditions). When present, cues were 100% valid in the sense that the probe always appeared at a cued location. Each participant completed seven (N = 1), 8 (N = 2), ten (N = 36), or 11 (N = 6) blocks of 60 trials as time permitted (participants were given a maximum of 1.5 h to complete the experiment). Performance feedback in the form of average absolute report error was given at the end of each block.

Fig. 2
figure 2

Hemifield Effects during Cue-two Trials in Experiment 1. We sorted participants average absolute recall errors (A), recall precision (B), swap rates (C), and guess rates (D) during cue-two trials according to whether the cued items appeared in the same visual hemifield or in different visual hemifields (“Diff”). Cue arrangement had no effect on any of these parameters. Error bars depict the 95% confidence interval of the mean

Full size image

Experiment 2: Endogenous orientation recall

Experiment 1 used sudden-onset cues that are known to trigger reflexive shifts of attention (Jonides & Yantis, 1988), thus Experiment 2 was conducted to examine whether the findings of Experiment 1 would generalize to a scenario where participants were encouraged to endogenously select cued items. Experiment 2 was identical to Experiment 1, with the exception that (a) we eliminated the cue-zero condition, and (b) we replaced the peripheral, exogenous cues used in Experiment 1 with central, endogenous cues. Specifically, we replaced the central fixation point used in Experiment 1 with a four-spoke fixation grid, where each spoke pointed towards one of the four stimulus locations (see Fig. 3A). Participants were retrospectively cued to remembered stimuli by changing individual spokes on the fixation grid from black to red for 100 ms (see Fig. 3A, which depicts an example cue-two trial). During cue-one trials we randomly cued one of the four stimuli, subject to the constraint that each location was cued equally often within a single block of 60 trials. During cue-two trials we randomly cued two of the four stimuli. Again, we made no attempt to control the spatial relationship between the cued stimuli (e.g., same vs. different hemifields), but examined whether this factor influenced performance in post hoc analyses (Fig. 4). Each participant completed five (N = 2), seven (N = 2), or eight (N = 24) blocks of 60 trials as time permitted (participants were given a maximum of 1.5 h to complete the experiment). Performance feedback (average absolute report error relative to the probed item) was given at the end of each block.

Fig. 3
figure 3

Design and Results of Experiment 2. A Task schematic showing a cue-two trial. Displays have been enlarged for exposition; see Methods for exact parameters. B-E Average absolute recall error (B), estimated recall precision (C), swap rates (D), and guess rates (E) as a function of cue condition. Error bars depict the 95% confidence interval of the mean

Full size image
Fig. 4
figure 4

Hemifield Effects during Cue-two Trials in Experiment 2. We sorted participants average absolute recall errors (A), recall precision (B), swap rates (C), and guess rates (D) during cue-two trials according to whether the cued items appeared in the same visual hemifield or in different visual hemifields (“Diff”). Cue arrangement had no effect on any of these parameters. Error bars depict the 95% confidence interval of the mean

Full size image

Experiment 3: Color recall

To examine whether the findings of Experiments 1 and 2 generalized to a new feature space, we conducted a third experiment where participants were retrospectively cued to one, two, or all four of the remembered colors (Fig. 5A). Cues were rendered by displaying an outline of the relevant stimuli location for 300 ms (see Fig. 5A). Stimulus colors were randomly selected from a 360° isoluminant CIE L*a*b color space with a minimum spacing of 30° (L = 70, a = -6, b = 14, radius 49). A sample display containing four colored squares (subtending 1° at a radial distance of 6° from fixation from a viewing distance of 57 cm) was presented for 500 ms followed by an 800-ms blank delay. A cue display was presented for 300 ms, followed by another 800-ms blank delay. Finally, a probe display containing one outline square was presented along with a color wheel; participants indicated their memory for the color that appeared at the outline location by clicking on the color wheel. Participants were instructed to prioritize accuracy, and no response deadline was imposed. During cue-one trials we randomly cued one of the four stimuli; there was no formal constraint on the number of times a location could be cued in a given block of trials. During cue-two trials we randomly cued two of the four stimuli. We made no attempt to control the spatial relationship between the cued stimuli (e.g., same vs. different hemifields). Participants completed 100 trials in the cue-four, cue-one, and cue-two conditions; breaks were given every 25 trials. Performance feedback was not given. Twenty of 25 participants additionally completed 200 trials of an unreliable cue-one condition in which the cue was informative but not always valid (the cued item was probed 50% of the time); the results are not analyzed here.

Fig. 5
figure 5

Design and Results of Experiment 3. A Task schematic showing a cue-one trial. Displays have been enlarged for exposition; see Methods for exact parameters. B-E Average absolute recall error (B), estimated recall precision (C), swap rates (D), and guess rates (E) as a function of cue condition. Error bars depict the 95% confidence interval of the mean

Full size image

Data analysis and statistics

Data from each experiment were analyzed using two complementary methods. To get an overall view of participants’ task performance, we computed estimates of average absolute recall error (i.e., the average absolute angular difference between the orientation or color reported by the participant and the actual probed orientation or color). We also fit participants’ recall data using a parametric model of the form:

$$p\left(\hat{\theta}\right)=\left(1-\gamma -\beta \right){\varPhi}_{\sigma}\left(\hat{\theta}-\theta \right)+\gamma \frac{1}{2\pi }+\beta \frac{1}{m}\sum_i^m{\varPhi}_{\sigma}\left(\hat{\theta}-{\theta}_i^{\ast}\right)$$
(1)

where θ is the target feature value, \(\hat{\theta}\) is the reported feature value, γ is the proportion of trials where the subject guesses, β is the probability of misremembering the target location, \(\left\{{\theta}_1^{\ast },\kern0.5em {\theta}_2^{\ast },\dots {\theta}_m^{\ast}\right\}\) are the values of the m nontarget items, and Φσ is a von Mises distribution with mean 0 and standard deviation σ (Bays et al., 2009). Fitting was performed in MATLAB using the open-source Analogue Report Toolbox developed by Paul Bays and colleague (available for download at http://www.bayslab.org/toolbox/index.php). This method returns participant- and condition-level maximum likelihood estimates for β, γ, and σ given inputs θ, \(\hat{\theta}\), and \(\left\{{\theta}_1^{\ast },\kern0.5em {\theta}_2^{\ast },\dots {\theta}_m^{\ast}\right\}\) over a range of seed values to account for local minima.

The effects of cue number (i.e., cue-one, cue-two, etc.) on β, γ, and σ were estimated via one-way repeated-measures analysis of variance (ANOVA) with cue number as the sole model factor. Where appropriate, false-discovery-rate-corrected post hoc comparisons were performed via repeated-measures t-tests. Throughout this article, we report condition averages, 95% confidence intervals, and effect sizes (η2 and Cohen’s d). Direct comparisons between cue conditions (e.g., cue-one vs. cue-two) deemed non-significant via frequentist analysis were further probed with Bayesian pairwise t-tests to quantify evidence favoring the null hypothesis. Bayesian analyses were run with uninformative (Jefferys) priors since prior research on the possibility of multiple simultaneous retrospective cue benefits is mixed (see Introduction). Bayesian analyses were performed using an open-source MATLAB toolbox (available for download at https://github.com/klabhub/bayesFactor). The result of a Bayesian t-test is a Bayes factor, typically denoted BF10. For example, a Bayes factor of 3.0 provides 3-to-1 odds favoring the alternative over the null hypothesis. Since Bayesian analyses were restricted to null effects (estimated using frequentist statistics), we computed an inverse Bayes factor BF01 describing the strength of evidence favoring the null over the alternative hypothesis, i.e., \({BF}_{01}=\frac{1}{BF_{10}}\). Finally, we evaluated model fits by randomly selecting and plotting histograms of recall errors with the best fitting model (Eq. 1) overlaid (see Figs. S1–S3 in the Online Supplementary Material (OSM)). While model fits varied from participant to participant and condition to condition, in general they provided a good approximation of participants’ recall errors.

Results

Experiment 1

The results of Experiment 1 are summarized in Fig. 1B–E. A one-way ANOVA applied to participants’ recall errors (Fig. 1B) revealed a main effect of cue number (i.e., 0, 1, 2, or 4), F(2, 72) = 22.81, p < 1e5, η2 = 0.388, and post hoc comparisons revealed that this effect was driven by superior performance during the cue-one relative to cue-two trials (M = 52.88° vs. 59.37°, respectively; t(36) = 5.456, p < 1e5, d = 0.34; 95% CI of the difference = 4.25–8.90°) and during cue-one relative to neutral trials (M = 52.88° vs. 59.40°; t(36) = 5.19, p < 1e5, d = 0.35; 95% CI of the difference = 4.25–9.08°). Recall performance during cue-two and neutral trials was statistically indistinguishable, t(36) = 0.04, p = 0.963; BF01 = 5.65 (for reference, a BF01 of 5.0 indicates 5-to-1 odds favoring the null hypothesis; see Data analysis and statistics, in the Methods section). These findings are consistent with earlier studies suggesting that location-based access to the contents of WM is limited to a single item at a time (e.g., Makovski & Jiang, 2007).

Next, we fit participants’ report errors with a parametric model (Eq. 1; Bays et al., 2009), which assumes that on a given trial participants (a) report the orientation of the probed stimulus with precision σ, (b) report the orientation of a non-probed stimulus with precision σ, or (c) randomly guess. Thus, we obtained participant- and cue-condition-level estimates of precision, the frequency of non-target reports (“swap errors”), and the frequency of random guessing. Parameter estimates obtained using this method are summarized in Fig. 1C–E. We quantified the effects of cue condition on memory precision, guess rates, and swap rates via independent one-way repeated-measures analyses of variance (ANOVAs). Cue type had no effect on estimates of recall precision (Fig. 1C; F(2, 72) = 2.64, p = 0.078) or swap rates (Fig. 1D; F(2, 72) = 2.90, p = 0.06). However, cue type had a significant effect on guess rates (Fig. 1E; F(2, 72) = 3.151, p = 0.048, η2 = 0.08). Visual inspection of Fig. 1E suggests that this effect was driven by lower guessing rates during cue-one relative to cue-two trials (M = 0.417 and 0.482, respectively; 95% CI of the difference = -0.023–0.136) and/or lower guessing rates during cue-one relative to neutral trials (M = 0.417 vs. 0.479; 95% CI of the difference = 0.015–0.122). However, post hoc comparisons between these conditions did not survive correction for multiple comparisons (t(36) = 1.86, 0.15, and 2.26; p = 0.107, 0.885, and 0.09; and BF01 = 1.20, 5.60, and 0.60 for the comparisons of cue-one vs. cue-two, cue two vs. neutral, and cue-one vs. neutral trials, respectively). Thus, Experiment 1 revealed a significant reduction in random guessing during cue-one relative to cue-two and uninformative cue trials (i.e., cue-zero or cue-four), but no difference in memory precision, random guessing, or swap errors across cue-two and neutral cue conditions.

We also investigated whether the spatial positions of the cued stimuli during cue-two trials influenced memory performance. For example, Delvenne and Holt (2012) found multiple simultaneous retrocue benefits when cued stimuli were arranged in different visual hemifields, but not in the same visual hemifield. We tested this possibility by sorting cue-two trials according to the spatial arrangement of retrospectively cued items (i.e., same vs. different hemifields) and recomputing average absolute recall error, recall precision, swap rates, and guess rates within each group. The results of this comparison must be treated with caution as our study was not explicitly designed to detect hemifield differences (e.g., we made no effort to ensure an equal number of same- vs. different-hemifield trials). Nevertheless, we found no evidence suggesting that cue arrangement had an effect on average absolute recall errors (Fig. 2A; t(36) = 1.272, p = 0.212; BF01 = 2.69), recall precision (Fig. 2B; t(36) = 1.098, p = 0.279; BF01 = 3.24), swap rates (Fig. 2C; t(36) = 0.555, p = 0.582; BF01 = 4.89), or guess rates (Fig. 2D; t(36) = 0.671, p = 0.507; BF01 = 4.58). The results of these comparisons provide modest evidence against the hypothesis that performance benefits during cue-two versus neutral trials (Fig. 1C–E) were obscured by the spatial arrangement of cue locations (i.e., same vs. different hemifields).

Experiment 2

The results of Experiment 1 demonstrated that participants��� absolute recall performance was worse during cue-two and neutral trials compared to cue-one trials. Furthermore, precision estimates, swap rates, and guess rates were statistically indistinguishable across cue-two and neutral trials. Next, we sought to replicate and extend these findings by examining whether a similar pattern was observed when participants were endogenously (as opposed to exogenously) simultaneously cued to prioritize multiple items. The results of Experiment 2 are summarized in Fig. 3B–E.

A one-way ANOVA applied to participants’ recall errors (Fig. 3B) revealed a main effect of cue type (i.e., cue-one, cue-two, cue-all), F(2, 54) = 13.81, p < 1e5, η2 = 0.388, and post hoc comparisons revealed that this effect was driven by superior performance during the cue-one relative to cue-two trials (M = 38.52° and 43.45°, respectively; t(27) = 4.89, p < 1e4, d = 0.279; 95% CI of the difference = 2.93–6.82°) as well as superior performance during cue-one relative to neutral trials (M = 38.52° and 42.09°, respectively; t(27) = 3.49, p = 0.002, d = 0.19; 95% CI of the difference = 1.54–5.50°). These results constitute a partial replication and extension of Experiment 1 and prior research (e.g., Makovski & Jiang, 2007). Cue type had no effect on recall precision (Fig. 3C; F(2, 54) = 1.205, p = 0.308) or swap rates (Fig. 3D; F(2, 54) = 0.595, p = 0.553), but did have a significant effect on guess rates (Fig. 3E; F(2, 54) = 3.580, p = 0.035, η2 = 0.12). Visual inspection of Fig. 3E suggests that this effect was driven by lower guessing rates during cue-one versus cue-two and cue-one versus neutral trials (M = 0.237, 0.294, and 0.256, for cue-one, cue-two, and neutral trials, respectively). Indeed, false discovery rate-corrected post hoc comparisons revealed significantly higher guess rates during cue-one versus cue-two trials (t(27) = 2.872, p = 0.0235, d = 0.276; 95% CI of the difference = 0.018–0.095), but no difference between cue-one and neutral trials (t(27) = -0.80, p = 0.431; BF01 = 3.72) or between cue-two and neutral trials (t(27) = 1.76, p = 0.135; BF01 = 1.28). The spatial positions of cued stimuli during cue-two trials (i.e., same vs. different hemifield) had no impact on participants’ recall errors (Fig. 4A; t(27) = 0.232, p = 0.812; 4.86), recall precision (Fig. 4B; t(27) = 1.511, p = 0.142; BF01 = 1.80), swap rates (Fig. 4C; t(27) = 0.550, p = 0.587; BF01 = 4.34), or guess rates (Fig. 4D; t(27) = 0.377, p = 0.709; BF01 = 4.67). Thus, the results of Experiment 2 are consistent with those of Experiment 1: first, participants’ absolute recall performance was worse during cue-two and neutral trials compared to cue-one trials, and this effect was driven by a reduction in guess rates during cue-one relative to cue-two trials (Fig. 3E). Second, estimates of recall precision, swap rates, and guess rates were statistically indistinguishable during cue-two relative to neutral trials.

Experiment 3

The results of Experiment 3 are summarized in Fig. 5B–E. A one-way ANOVA applied to participants’ recall errors (Fig. 5B) revealed a main effect of cue number (F(2, 48) = 40.91, p < 1e5, η2 = 0.630). Post hoc comparisons revealed that this effect was driven by superior performance during cue-one relative to cue-two trials (M = 28.78° and 46.57°, respectively; t(24) = 8.07, p < 1e5, d = 1.107; 95% CI of the difference = 13.56–22.13°) and during cue-one relative to neutral trials (M = 28.78° and 47.41°, respectively; t(24) = 6.98, p < 1e5, d = 1.127; 95% CI of the difference = 13.73–24.17°). Cue number also had a significant effect on swap rates (Fig. 5D; F(2, 48) = 8.921, p = 0.0005, η2 = 0.271) and guess rates (Fig. 5E; F(2, 48) = 5.273, p = 0.009, η2 = 0.180). Post hoc analyses revealed significantly greater swap rates during cue-two versus cue-one trials (M = 0.113 vs. 0.025, respectively; t(24) = 3.99, p = 0.0016, d = 0.797; 95% CI of the difference = 0.047–0.133) and during neutral vs. cue-one trials (M = 0.066 vs. 0.025, respectively; t(24) = 3.026, p = 0.0087, d = 0.838; 95% CI of the difference = 0.014-0.067), but no difference in swap rates during cue-two and neutral trials (t(24) = 1.86, p = 0.075; BF01 = 1.07). Complementary analyses revealed significantly greater guess rates during neutral relative to cue-one trials (M = 0.305 vs. 0.168, respectively; t(24) = 3.27, p = 0.009, d = 0.679; 95% CI of the difference = 0.061–0.221), but no difference in guess rates during neutral relative to cue-two trials (M = 0.305 vs. 0.270, respectively; t(24) = 0.847, p = 0.405; 95% CI of the difference = -0.044–0.114; BF01 = 3.43). Differences in guess rates during cue-one and cue-two trials were not statistically significant, (t(24) = 2.12, p = 0.066; 95% CI of the difference = -0.002-0.188; BF01 = 0.704). Finally, we found no effects of hemifield (i.e., same vs. different) on task performance during cue-two trials (Fig. 6; t(24) = 1.262, 1.274, 0.392, and 0.276 for recall errors, recall precision, swap rates, and guess rates, respectively; all p > 0.215). These conclusions were supported by Bayesian t-tests (BF01 = 2.33, 2.30, 4.42 and 4.58 for recall errors, recall precision, swap rates, and guess rates, respectively). Thus, the results of Experiment 3 are consistent with those of Experiments 1 and 2, with the exception that swap errors were more common during cue-two and cue-four relative to cue-one trials (Fig. 5D). However, once again there were no observed differences in recall precision or error types (i.e., swaps vs. guesses) during cue-two relative to neutral trials.

Fig. 6
figure 6

Hemifield Effects during Cue-two Trials in Experiment 3. We sorted participants average absolute recall errors (A), recall precision (B), swap rates (C), and guess rates (D) during cue-two trials according to whether the cued items appeared in the same visual hemifield or in different visual hemifields (“Diff”). Cue arrangement had no effect on any of these parameters. Error bars depict the 95% confidence interval of the mean

Full size image

Discussion

Retrospective cue paradigms can be used to study the consequences of allocating attention to stimuli stored in working memory (Griffin & Nobre, 2003; Landman et al., 2003). Spatial cues instructing participants to retrospectively prioritize a single item stored in memory leads to a decrease in recall errors, an increase memory precision, a decrease in the probability of reporting a non-target item from memory (i.e., a swap error), and a decrease the probability of randomly guessing (Gunseli et al., 2015; Makovski & Pertzov, 2015; Murray et al., 2013; Pertzov et al., 2013; Souza et al., 2014, 2016). Sequentially presented retrospective cues also improve WM performance, either for the last cued item in a sequence (Landman et al., 2003; Li & Saiki, 2014; Maxcey et al., 2015) or for all cued items in a sequence (Souza et al., 2015), depending on the specific task structure. Thus, participants can access information stored in WM based on position information, and doing so improves memory performance. However, whether location-based access to the contents of memory extends to multiple items at the same time is less clear, with some findings arguing against this possibility (e.g., Makovski & Jiang, 2007; Oberauer & Bialkova, 2009) and others suggesting that it is possible only under certain circumstances (e.g., when simultaneously cued items appear in different visual hemifields; Delvenne & Holt, 2012). Most studies investigating this specific possibility have relied on discretized measures of WM performance, including change detection accuracy or average (absolute) recall error. It is possible that location-based access to WM content extends to multiple items at the same time, but that this improves some aspects of WM performance (e.g., an increase in memory precision) while worsening others (e.g., an increased likelihood of confusing two prioritized items). We tested this possibility using a parametric modeling approach (Bays et al., 2009). The results of three experiments revealed no evidence to support the conclusion that location-based access to the contents of WM can extend to more than one item at a time.

One earlier study reported that location-based access to WM content can extend to multiple items at the same time when cued stimuli appeared in different visual hemifields but not the same visual hemifield. We were unable to replicate this finding. However, we caution that our experiments were neither designed nor optimized to capture these effects. Specifically, the locations of retrospectively cued items were randomly selected during each cue-two trial. Since there are four possible different-hemifield cue combinations (i.e., upper and lower visual fields as in Fig. 1A, or across the diagonals as in Fig. 3A) and only two possible same-hemifield cue combinations (i.e., the two left or two right stimuli), the latter were under-represented in our analysis. This, in turn, may have obscured hemifield effects on location-based access to WM content.

The lack of a performance difference between cue-two relative to cue-four trials could reflect participants’ inability to use multiple spatial cues. We think this unlikely for several reasons. First, there is ample evidence showing that human observers can successfully use multiple simultaneously presented cues to allocate attention in the external environment (e.g., Awh & Pashler, 2000; Ester et al., 2014; Franconeri et al., 2007; Müller et al., 2003) and to gate access to WM (e.g., Allen & Ueno, 2018; Makovski & Jiang, 2007). Second, several studies have documented improved WM performance following the presentation of multiple sequentially presented cues (e.g., Li & Saiki, 2014; Maxcey et al., 2015; Souza et al., 2015). Thus, the limiting factor that precludes performance benefits following multiple simultaneously presented spatial cues must involve selecting and prioritizing the appropriate items already stored in memory, rather than processing or interpreting the retrospective cue.

Simultaneous cue benefits on WM performance may vary depending on task demands. Although the current study found no benefit for multiple simultaneously presented spatial retrocues, Allen and Ueno (2018) reported that multiple simultaneously presented pre-cues (i.e., presented prior to the onset of a memory encoding display) improved performance on a WM discrimination task. This raises the intriguing possibility that the contents of memory can be assigned different levels of priority during WM encoding or consolidation, but not after encoding or consolidation is complete. Future research could test this possibility by directly comparing the effects of multiple simultaneously presented pre- and retro-cues in the same participants cohort.

Single location-based access to WM content can be contrasted with feature-based access to WM content that extends to multiple items. For example, Arnicane and Souza (2021) asked participants to memorize arrays containing two blue and two red triangles, then presented participants with a color cue indicating that one item from the set of cue-matching items (i.e., blue) would be probed for report at the end of the trial. Here, robust retrocue benefits were observed when the array of to-be-remembered information contained one cue-matching stimulus or two cue-matching stimuli. Thus, feature cues seem to enable access to multiple remembered stimuli at the same time. This result aligns with psychological and neurophysiological results suggesting that directing attention to a specific feature enhances sensory processing of that feature throughout the visual field (e.g., Martinez-Trujillo & Treue, 2004; Saenz et al., 2002; Serences & Boynton, 2007), or that the identity of a remembered feature can be decoded from brain areas that are not retinotopically mapped to the position where that feature appeared (Ester et al., 2009; Ester et al., 2015), and indicates that different limitations govern location- and feature-based access to WM content.

To summarize, we used a parametric model to examine whether location-based access to WM content can extend to multiple remembered objects at the same time. While we found clear evidence for a memory performance advantage when participants were instructed to retrospectively prioritize information that appeared in a single location, this performance benefit disappeared when participants were instructed to retrospectively prioritize information that appeared in different locations. Thus, our findings support the conclusion that location-based access to WM content is limited to a single item at a time.