Human relational memory requires time and sleep

Abstract

Relational memory, the flexible ability to generalize across existing stores of information, is a fundamental property of human cognition. Little is known, however, about how and when this inferential knowledge emerges. Here, we test the hypothesis that human relational memory develops during offline time periods. Fifty-six participants initially learned five "premise pairs" (A>B, B>C, C>D, D>E, and E>F). Unknown to subjects, the pairs contained an embedded hierarchy (A>B>C>D>E>F). Following an offline delay of either 20 min, 12 hr (wake or sleep), or 24 hr, knowledge of the hierarchy was tested by examining inferential judgments for novel "inference pairs" (B>D, C>E, and B>E). Despite all groups achieving near-identical premise pair retention after the offline delay (all groups, >85%; the building blocks of the hierarchy), a striking dissociation was evident in the ability to make relational inference judgments: the 20-min group showed no evidence of inferential ability (52%), whereas the 12- and 24-hr groups displayed highly significant relational memory developments (inference ability of both groups, >75%; P < 0.001). Moreover, if the 12-hr period contained sleep, an additional boost to relational memory was seen for the most distant inferential judgment (the B>E pair; sleep = 93%, wake = 69%, P = 0.03). Interestingly, despite this increase in performance, the sleep benefit was not associated with an increase in subjective confidence for these judgments. Together, these findings demonstrate that human relational memory develops during offline time delays. Furthermore, sleep appears to preferentially facilitate this process by enhancing hierarchical memory binding, thereby allowing superior performance for the more distant inferential judgments, a benefit that may operate below the level of conscious awareness.

Conflict of interest statement

The authors declare no conflict of interest.

Figures

Fig. 1.

The transitive inference task and experimental paradigm. (A) Six visual object stimuli (see SI Fig. 4, illustrated conceptually here as A–F) were combined as five premises (premise pairs), where “>” describes the relationship “should be selected over.” Unknown to subjects, these premise pairs formed an ordered hierarchy such that A>B>C>D>E>F. (B) To evaluate knowledge of the hierarchy, subjects were later tested using the premise pairs and novel “inference” pairs not previously learned. These inference pairs involved either one degree (B-D and C-E) or two degrees (B-E) of item separation. (C) All participants initially learned the premise pairs during a training session, involving reinforcement cues signifying which item was correct. Immediately following learning, the reinforcement cues were removed, and subjects were tested on the premise pairs to measure the extent of learning without feedback. Following a delayed offline time interval of 20 min, 12 hr, or 24 hr, subjects were again tested on the premise pairs but were also probed for hierarchical knowledge by using the novel inference pairs.

Fig. 2.

Delayed test performance (percent correct) across groups. (A) Premise pair retention following the three delayed time intervals/groups (averaged across A-B, B-C, C-D, D-E, and E-F pairs; individual pair values provided in Table 1). Irrespective of the time delay group, all subjects expressed near-identical premise pair knowledge following the offline period. (B) Inference pair performance at the delayed test session across the three groups (averaged across all novel inference pairs, B-D, C-E, and B-E). There was no evidence for the development of hierarchical knowledge after the short (20-min) delay, with inference performance not significantly different than chance. In contrast, following time delays of either 12 or 24 hr, significantly above-chance performance was evident, with accuracy scores significantly different than the 20-min group. Performance between the 12- and 24-hr groups was not significantly different. (C) Inference pair performance also at the later delayed test session, across the three groups, but separated according to the distance of item separation, one degree of item separation (averaged B-D and C-E pairs) or two degrees (B-E pair). Asterisks represent significant performance difference (P < 0.05). Error bars represent standard errors.

Fig. 3.

Delayed inference performance and corresponding confidence ratings. (A) Delayed inference pair performance (percent correct) across the 12- and 24-hr delay. Twelve-hr groups are separated according to the Wake and Sleep subgroup assignment, and inference performance is separated according to the distance of item separation, one degree of item separation (averaged B-D and C-E pairs) or two (B-E pair; individual pair values provided in Table 1). In the Wake group, performance was not different across the one- and two-degree inference judgments. However, in both groups that experienced an intervening night of sleep (e.g., 12-hr Sleep and 24-hr groups), significantly better performance was expressed on the more distant two-degree inference judgment compared with the one-degree judgment. (B) Corresponding confidence ratings (from 1 to 7, with 7 representing the highest confidence) for the one- and two-degree inference pairs in the 12-hr Wake and Sleep subgroups, as well as the 24-hr group. In contrast to the marked performance differences in performance for the one- and two-degree inference in the 12-hr Sleep and 24-hr groups, no corresponding increase in subjective confidence was apparent, in any group. Box-and-whisker plots describing individual subjects' distributions are provided in SI Fig. 5. Mean subjective confidence ratings for the one- and two-degree inference pairs in the 20-min group were 3.7 and 3.4, respectively. Asterisks represents significant performance difference (P < 0.05). Error bars represent standard errors.