Long-term memory guides resource allocation in working memory

The study investigates how prior knowledge stored in long-term memory (LTM) can guide the allocation of limited working memory (WM) resources. WM supports goal-directed behavior but has severe capacity limits, necessitating selective encoding and prioritization of relevant information. Prior work shows that prioritization improves WM performance and that both behavioral relevance and uncertainty guide resource allocation. LTM can facilitate WM via compression/chunking and by guiding attention toward probable locations or features. The authors hypothesized that when one item’s likely location is known in advance, participants would strategically deprioritize encoding that familiar item in WM and instead prioritize encoding novel items, thereby improving overall memory performance. They further predicted potential costs for novel items appearing near the familiar item’s expected location due to attentional deprioritization.

Prior research demonstrates flexible and unequal distribution of WM resources, with attention prioritizing relevant items or features and performance better explained by prioritization than set size. Information currently in WM can capture attention and influence behavior, with uncertainty guiding decisions and gaze allocation. LTM supports WM through compression/chunking, enabling efficient representations and faster processing for familiar items, and can guide attention to probable locations without interference when relevant. People can learn environmental statistics implicitly, but the use of priors may be contingent on awareness for compression-like benefits. Together, this literature supports the possibility that explicit prior knowledge can be used to strategically shift attention and WM resources toward novel, unpredictable information while relying on LTM to handle familiar, predictable aspects.

Design: Full-report delayed-estimation visual working memory task with six colored dots per trial. One color (the familiar item) appeared at locations sampled from a von Mises distribution centered on a participant-specific mean with SD≈20°; the other five colors (novel items) appeared uniformly around a circle. Participants reported locations for all six colors in any order each trial. Participants: 62 subjects (41 female; ages 18–21) with normal or corrected vision, not colorblind, provided informed consent. IRB-approved (University of Texas at Austin; protocol #2013-10-0110). Compensation via course credit or $12/h (n=6). Apparatus/Stimuli: Implemented in Python/PsychoPy on a 21.5-inch iMac (1920×1080, 60 Hz). Six HSV colors (red, orange, yellow, green, cyan, blue; fixed hues) appeared at 3.25° eccentricity with at least 30° angular separation. A visual mask (random noise ring) followed encoding. During response, six colored dots appeared at the bottom to select which color to report, then participants clicked the corresponding location on a response ring. Task timing per trial: Encoding 1000 ms; mask 200 ms; delay 750 ms; full-report (six responses, any order) until completion; feedback displayed true and reported locations plus per-response points (Gaussian scoring centered at 0°, SD=30°, max 10 points per item). Blocks ended when a point goal (calibrated from practice to yield ~20 trials per block) was reached. Total 8 blocks (mean ≈18.43 trials per block). Prior information: Participants learned the color and mean location of the familiar item via an explicit training phase showing a colored arc indicating ±2 SD (80°) around the prior center. After ~3 practice trials, they had to recall color and location (within the 80° arc) to proceed. This prior information then remained constant for the experiment without further on-screen reminders. Experiments: Two groups differed only in the first block. Experiment 1 (n=39): first block had no prior information; all items were novel (uniform). From block 2 onward, one familiar item with prior and five novel items. Experiment 2 (n=23): prior information provided before block 1; all blocks had one familiar and five novel items. Measures and analyses: Primary precision metric was mean resultant vector length (MRVL; 0–1). Error was angular difference between response and true location. Statistical tests used Pingouin in Python; two-sided tests, Bonferroni-corrected where applicable; bootstrapped 95% CIs in figures. Analyses included within- and between-subject t-tests, repeated-measures two-way ANOVA (item type × response number), linear regressions of precision on response number, chi-square goodness-of-fit for response order of the familiar item, binomial tests for early vs late reporting, and location-based ANOVAs and regressions to assess spatial biases. Where noted, degrees of freedom varied slightly when some participants lacked familiar-item responses at certain positions.

- Benefit of prior information: In Experiment 1, introducing prior information significantly increased precision (MRVL) for both the familiar item and the novel items compared to the no-prior first block: familiar t(38)=10.81, p(Bon)<0.0001, d=2.44; novel t(38)=8.59, p(Bon)<0.0001, d=1.09. The increase was greater for the familiar item: t(38)=6.34, p(Bon)<0.0001, d=1.28. - Practice control: Experiment 2 showed a practice effect (first block < remaining blocks; t(22)=2.48, p=0.021, d=0.48). Comparing first blocks across experiments (with-prior vs without-prior) still showed higher precision with prior: t(60)=3.78, p=0.0004, d=0.99. - Response number and item type: Significant main effects of item type (F(1,61)=94.63, p<0.0001, ηp²=0.61) and response number (F(5,305)=144.25, p<0.0001, ηp²=0.70) with interaction (F(5,305)=41.08, p<0.0001, ηp²=0.40). Familiar vs novel differed in the last three response positions: response 4 t(60)=7.49, d=1.14; response 5 t(59)=11.71, d=1.95; response 6 t(60)=9.38, all p(Bon)<0.0001. Precision decreased with later responses: familiar β=-0.024 (p<0.0001, adj R²=0.06); novel β=-0.087 (p<0.0001, adj R²=0.58). - Response order bias for familiar item: Reporting position of the familiar item was non-uniform (χ²(5,N=8351)=379.58, p<0.0001); more likely in positions 4–6 than 1–3 (57.1%, binomial p<0.0001). Novel items reported before the familiar item had higher precision than those reported after (t(61)=11.28, p<0.0001, d=1.60). Position-specific effects for responses 2–5: response 2 t(59)=2.88, p(Bon)=0.0056, d=0.50; response 3 t(61)=2.47, p(Bon)=0.016 (ns after correction), d=0.35; response 4 t(60)=4.98, p(Bon)<0.0001, d=0.64; response 5 t(60)=6.08, p(Bon)<0.0001, d=0.88. This analysis showed an experiment-group effect (F(1,60)=5.48, p=0.023, ηp²=0.08). - Spatial effects relative to prior center: Location influenced precision for both familiar (F(9,441)=2.52, p=0.0051, ηp²=0.05) and novel items (F(35,2135)=3.72, p<0.0001, ηp²=0.06). Novel items near the prior center (-10° to 10°) had worse precision than elsewhere (t(61)=4.54, p<0.0001, d=0.78). Familiar items near the center had improved precision (t(61)=10.17, p<0.0001, d=1.08). Familiar-item errors showed a bias toward the prior mean: error became 3.22° closer to the mean per 10° bin from center (β=-3.22, p<0.0001; ANOVA on errors F(9,441)=12.72, p<0.0001, ηp²=0.20). Novel-item errors showed no location-dependent bias (F(9,540)=1.10, p=0.36). No main effects of experiment group for these spatial bias analyses, though an interaction with location emerged (F(9,540)=3.29, p=0.0006, ηp²=0.05), with a location effect present in Experiment 2 only (F(9,198)=2.54, p=0.0088, ηp²=0.10).

Findings demonstrate that explicit prior knowledge about one item’s probable location improves overall WM performance in a full-report task: precision increased for both familiar and novel items when prior information was available. Behavioral patterns suggest a strategic allocation of WM resources: participants tended to rely on LTM to report the familiar item, freeing WM resources to encode novel items with higher precision. Response precision decreased with later responses, and participants often reported the familiar item in the latter half of the sequence. Novel items reported before the familiar item were more precise than those reported after, aligning with a strategy where the familiar item serves as a boundary between higher-precision (prioritized) and lower-precision (unprioritized/guess-like) responses. Spatial analyses showed reduced precision for novel items near the familiar item’s prior center and a response bias toward the prior mean for familiar items, indicating an attentional deprioritization of the prior region during encoding and reliance on prior knowledge during recall for the familiar item. Together, results support a dynamic interaction between WM and LTM, where prior knowledge guides attention away from predictable information and toward novel, unpredictable stimuli to optimize limited WM resources.

The study shows that people use long-term memory to guide the strategic allocation of limited working memory resources. When one item’s likely location is known, participants deprioritize encoding that familiar item and allocate attention to novel items, improving overall report precision in a capacity-limited scenario. Full-report patterns and spatial effects corroborate this strategy. This work advances understanding of WM-LTM interactions and the role of prior knowledge in optimizing memory performance. Future research should test conditions with variable or unreliable priors, include confidence measures to examine metacognitive strategies, collect reaction times to assess potential compression/recoding processes, and examine whether implicit learning of priors yields similar resource-allocation benefits.

- Generalizability may be limited to contexts with stable, explicitly learned priors; participants always knew the familiar item would appear and where it tended to be. - Potential reliance on explicit awareness; compression-like strategies may not occur with implicit priors. - No confidence ratings were collected, limiting inferences about metacognitive ordering strategies. - Reaction times in the full-report paradigm were not informative, constraining tests of compression/recoding accounts. - Some analyses showed modest experiment-group interactions, indicating procedural differences can influence certain effects.