Information encountered several times gets remembered as having first arrived earlier than information encountered once, even when the repeated material is the better remembered of the two. That is the finding of six experiments by Brynn Sherman of the University of Pennsylvania and Sami Yousif of the University of North Carolina at Chapel Hill, published in Psychological Science in April 2025. They call the pattern the temporal repetition effect.
The sharpest demonstration used a forced choice. Fifty adults watched a long stream of object images, then saw two of those objects side by side and were asked which one they had seen first. One of the pair had recurred through the sequence. The other had appeared exactly once, and always slightly earlier. Participants chose the repeated object 78 percent of the time. On every trial, the correct answer was the other object.
A sequence engineered so the fillers always came first
Participants were recruited through the online platform Prolific. They watched images of objects appear one at a time, each on screen for 1.5 seconds, and judged whether each was larger or smaller than a shoebox. In the first experiment nobody was told their memory for timing would be tested.
Underneath, the sequence was built to make a clean comparison possible. Some objects, the targets, recurred across blocks. Others, the fillers, appeared once. Each target was paired with the filler that had appeared immediately before its first showing, so any comparison involved two objects encountered within seconds of each other. The design handed the fillers a genuine advantage: they really did come first.
Participants then placed each object on an on-screen timeline spanning the encoding phase, and repeated objects landed further back than their matched fillers. The paper’s phrasing for that gap is worth keeping intact: repeated items were remembered as much as 43 percent further back along the timeline than items that were not repeated. The paper sets that gap against the fillers rather than against the full length of the timeline, though it never spells the denominator out. Forty-three of the 50 participants showed the effect.
Something else happened alongside the distortion. Participants tracked the true running order slightly better for repeated objects than for single ones, so repetition appeared to sharpen the sense of sequence while dragging the apparent starting point backwards. That comparison was not preregistered, the effect was small, and in the week-long experiment the pattern ran the other way, though not to significance.
The repetitions people noticed did the work
More exposures generally meant a larger distortion, though the evidence for that is uneven rather than uniform. In the first experiment the scaling was clear, and objects shown three times were misplaced further back than objects shown twice. The step from three exposures to five was much weaker, significant but small. In the second experiment, which had only 20 participants, the scaling did not appear at all, and the authors attribute that, tentatively, to reduced statistical power.
A fourth experiment got closer to a mechanism. Participants were asked during encoding whether they had seen each object before, which recorded what people actually registered rather than what the schedule dictated. The number of repetitions a participant remembered predicted the distortion, while the scheduled repetition structure added nothing once remembered repetitions were in the model. The authors’ conclusion carries a hedge that matters: the effect is “in part” dependent on recalling that an image had been presented before, rather than on mere repetition. In the week-long experiment the same model found the true repetition structure did improve the fit after all, so the mechanism is not settled by these data.
Two alternative explanations were tested and neither survived intact. Because the first experiment separated repeats with pauses, the effect might have come from those event boundaries; a version using one uninterrupted stream of 250 images produced it anyway, and the authors conclude it “does not seem to critically depend” on explicit boundaries. Participants might also have been reasoning it through, on the theory that anything seen often must have started early. A supplementary experiment asked how much they had leaned on that strategy: self-reported use averaged 39 percent of trials, but neither that figure nor a separate influence rating predicted the size of an individual’s distortion. The paper’s word for its own reading of this is “cautiously,” and it states that it does not fully reject a metamemory account.
A third experiment addressed the most obvious worry. Participants were told in advance they would be asked when they first saw each image. Knowing the question did not help, and the effect persisted at a similar magnitude, by the authors’ reading possibly a stronger one.
Across a working week
Streams of images run for minutes. To test the illusion at something nearer the scale of ordinary life, the researchers ran a version across five consecutive days, Monday to Friday, with the memory test on the following Monday. One hundred people were recruited; 60 completed all six days and met the response-rate threshold, and those 60 are the reported sample.
The effect replicated. Participants placed the first appearance of repeated objects further back than that of fillers, and 47 of the 60 showed the pattern.
Memory for timing otherwise largely collapsed. Only 30 of the 60 participants placed objects in better than random order, and at the group level the correlation between remembered and actual position was not reliably above zero. Recognition memory was considerably lower than in the shorter experiments, though still high. So the week-long result shows the illusion surviving inside a temporal memory that had become close to useless, which is a more limited finding than the illusion scaling smoothly up to real life.
The participants who were dropped first
Two of the exclusion criteria concern performance on the measure itself. Across the first four experiments, participants were removed not only for failing to finish or for skipping trials, but also if their timeline judgments were uncorrelated with the true order of the images, or if their hit and false-alarm rates fell outside set bounds. In the first experiment 31 people were removed to reach a usable sample of 50. The paper states that none of these exclusions relate to the effect being measured. It remains the case that people with poor memory for timing were screened out of the experiments measuring memory for timing, and that in the week-long version, where that screen was not applied, half the sample had no reliable sense of order at all.
The sample is the other stated limit, and the authors do not bury it. Every participant was an adult in the United States, recruited online. The paper’s generalizability statement restricts its claims to “educated, Western, English-speaking adults” and calls wider application an open question.
Then there is the headline number. The abstract describes the illusion as amounting to “as much as a 25% distortion in perceived time,” and that figure appears exactly once, with no statement anywhere of what it is 25 percent of. The discussion, which does give a denominator, puts it lower and narrower: “as much as 24% of the full length of the experiment,” specifying that this applies to the shorter studies. The 43 percent from the first experiment reads as that same gap set against the fillers instead, though the paper does not spell that out either. Anyone quoting a single figure for this effect should say which one, and what it is a percentage of.
The paper also marks out the boundary of its own claim. In a small pilot, participants asked when they had last seen an object showed the reverse pattern, remembering repeated objects as more recent. The effect described here is specific to the onset of an experience. It does not license the broader idea that repetition makes things feel older in general, and the authors note that the reversal fits existing work on recency judgments.
Theory has not caught up
The result runs against what the field would have expected, with a caveat the paper is careful to include. Established accounts hold that repeating something either creates additional time-stamped memory traces or strengthens the memory overall, and that either way the item should feel more recent. But those accounts were developed for recency judgments, and the paper says only that they “presumably” make the same prediction for judgments about when something first happened. It also observes that few theories make direct predictions about this at all. The pilot reversal suggests the conventional prediction may hold perfectly well for the question it was built for.
What remains is a well-documented anomaly no current framework comfortably accommodates, and the authors say the account they find most promising would require elaboration to do it. The experiments were preregistered, the data and analysis code are public, and the journal’s STAR team independently confirmed the results reproduce computationally. The transparency statement is also specific about where the preregistrations fell short: the exact statistical tests were not specified in advance, the target-versus-filler memory comparisons were not preregistered, some analyses were added in response to reviewers, and a design typo in the first preregistration carried through into most of the later ones.
Whether any of this matters outside a timeline task is the question the experiments do not reach. The authors raise the possibility that an event recalled as long ago gets imagined less concretely, and wonder whether that distance dulls the impulse to prepare for its return, offering the pandemic as their illustration. They present it as a question rather than a result, and no data in the paper speaks to it.
