Zeigarnik effect

Two separate predictions sit inside that one story, and they have come apart. The memory prediction became the Zeigarnik effect. The action prediction, tested by Maria Ovsiankina a year later, became the Ovsiankina effect: when people are interrupted and then left alone, they resume. Modern work does not support the tension account of the memory half, and the better-replicated resumption half does not need it. Ghibellini and Meier's own 2024 experiment points at a different mechanism for whatever memory advantage does appear: a discrepancy between expecting to succeed and then failing, which makes the unsolved item stand out, closer to a distinctiveness effect than to a charged tension system. For interface work, the useful version of the principle is the resumption tendency, not the memory advantage. People who left something half-done tend to take the opportunity to finish it. They do not reliably remember what it was, which is exactly why the interface has to carry the memory for them.

The original numbers were large and have never been matched. As tabulated by MacLeod (2020), Zeigarnik's Experiment 1 tested 32 adults on 22 tasks and produced an interrupted-to-completed recall ratio of 1.9 — nearly twice as many interrupted as completed tasks recalled — with 81.2% of participants recalling more interrupted than completed tasks. Her Experiments 2 to 4 gave ratios of 2.0, 1.9 and 2.1. MacLeod also reports Marrow (1938), who replicated the pattern on 30 participants with 20 paper-and-pencil tasks and got a ratio of 1.77.

Then the record turns. Butterfield's review concluded that the Zeigarnik effect is far from being the invariable result in the interrupted task paradigm, and that frequently more completed than incompleted tasks are recalled. Van Bergen set out in 1961 to reproduce Zeigarnik exactly, tested 34 participants, and found a ratio of .88, with 20 of them recalling more completed than interrupted tasks; she ran seven experiments in total and reported that in none of them was the phenomenon verified, and that of the 44 papers she reviewed as replication attempts, fewer than a third reported a Zeigarnik effect.

The 2025 meta-analysis settles the size of it. Ghibellini and Meier screened 1,349 de-duplicated publications down to 59, of which 38 investigated the Zeigarnik effect, 20 the Ovsiankina effect and one both. On the most commonly used measure, the ratio of mean interrupted recall to mean completed recall was 0.99 across 38 publications, and still 0.99 with Zeigarnik's own data removed: no difference. On the measure they recommend, interrupted tasks made up 49.16% of all tasks recalled, a slight disadvantage. Where an effect size could be computed at all, from eight publications, the weighted average was dz = 0.15. Zeigarnik's own preferred ratio, which averages per-participant ratios and is distorted by extreme values, gives 1.13 including her data and 1.09 without it, which is how the effect survived as folklore.

The resumption half holds up. In the same meta-analysis the weighted resumption rate was 67.00% across 21 publications, and 66.79% with Ovsiankina's own study excluded, against a 50% reference point. The authors are careful about this: resumption was studied descriptively, with heterogeneous procedures, so treat 67% as indicative rather than as an effect size. Their conclusion is that the Ovsiankina effect is a general tendency while the Zeigarnik effect lacks universal validity. A modern digital version exists: Birk, Mandryk and Baumann interrupted people playing a computer game with a fake network-interruption notice they could dismiss with one click or let time out, across three studies with 208, 457 and 210 participants, and found that willingness to resume tracked a self-regulation trait (demand-related action orientation) rather than being uniform, and that repeated exposure led to adaptation.

The best-evidenced neighbouring finding is about open goals occupying attention rather than memory. Masicampo and Baumeister had 73 undergraduates (69 analysed) write about either two important unfinished errands, the same errands plus a specific plan, or two recently completed tasks, then read 3,200 words of a novel. Self-reported intrusions during reading differed by condition, F(2, 68) = 4.67, p = .013, partial eta squared = .13: the unfinished group reported more distraction by task-related thoughts (M = 3.00, SD = 1.90) than controls (M = 1.82, SD = 1.44), F(1, 66) = 7.46, p = .01, while the plan group reported fewer than the unfinished group (M = 1.77, SD = 1.06), F(1, 66) = 6.41, p = .014, and did not differ from controls. In their Study 3, 126 participants primed with the unfulfilled goal of exercising showed higher accessibility of goal-related words than controls, F(1, 122) = 5.73, p = .018, but the effect was moderated by how much they valued the goal (interaction β = .21, t(83) = 2.02, p = .047; β = .48, t(83) = 3.37, p < .001 at high importance, nothing at low importance). Treat these as suggestive rather than settled: they are small single-lab samples from 2011, with no independent direct replication found.

Field evidence that open loops cost something is stronger on the harm side than the engagement side. Syrek, Weigelt, Peifer and Antoni matched 357 Friday and Monday observations from 59 employees over 12 weeks and found that unfinished tasks at the end of the week predicted impaired weekend sleep, mediated by affective rumination, with a separate between-person effect of accumulated unfinished tasks over three months.