z = 0.295 across 26 infants, confidence interval clear of zero, P = 0.008. The result holds with a continuous looking measure instead of a binary one (fig. S1). Controls address the standard confounds of this design: item memorability (fig. S2), serial position (fig. S3), and whether the brain signal tracks a general strength of preference rather than preference for the old item specifically (fig. S4). Trials were included only if the infant was looking.

The effect size is modest and the sample is 26. The paradigm is borrowed from adult cognitive neuroscience, where it has a long record, and the adaptation to infants (interleaved tests, looking time instead of a button press) is reasonable. Within the task as built, this is the firmest finding in the paper.

Averaged across infants, the proportion of test trials with more looking at the old picture was 0.510, against a chance value of 0.5 (P = 0.540). Neither age group differed from chance. Bootstrapping the older infants to a full sample of 26 still gave no preference (P = 0.398). The paper reports all of this plainly in the second results paragraph.

Everything downstream rests on sorting trials within each infant by a looking measure that, averaged over infants, could not tell old pictures from new ones. The subsequent memory design is built to use trial-to-trial variance, and the authors say so, but a reader of the abstract would not learn that the memory assay showed no memory on average. The paper's remedy is to split infants by their average preference. The hippocampal effect is strong in the half with higher preferences (P < 0.001) and not reliable in the other half (P = 0.276). That split selects on the outcome and was not described as planned. The authors also note that familiarity preferences have traditionally been read as a sign of partial or incomplete encoding, which they accept as a feature of what they measured.

The older 13 carry the effect (P < 0.001) and the younger 13 show none (P = 0.898), with a significant group difference and a continuous correlation with age (r = 0.385, P = 0.034). Six possible age confounds were partialed out (usable trials, attention to the screen, looking preferences, hippocampal volume and others, table S2). Signal variance was comparable across ages, and visual category areas responded comparably to faces, objects and scenes in both groups, which argues against the younger infants simply not looking.

Thirteen infants per group. The continuous correlation is P = 0.034 with that sample. The two age groups are 4 to 9 months and 12 to 24 months, so there are no infants in the 9 to 12 month window where the onset is placed. "Around one year" is a boundary drawn through a gap in the data. Stimulus duration was fixed at two seconds for every age; younger infants are known to need longer to encode, and the authors flag this and recommend infant-controlled habituation designs in future. The age effect was present in both anterior and posterior hippocampus, with the posterior numerically stronger, which the paper reports without a test.

Posterior hippocampus P = 0.002, anterior P = 0.140. The paper then states that the comparison between posterior and anterior "did not reach significance" (P = 0.118). The same pattern repeats in the high-preference subgroup: posterior P < 0.001, anterior P = 0.242, no direct test reported.

A significant result in one region and a null in another is not evidence that the regions differ. The direct comparison is the test, and it was null. The results text says this. The discussion and the abstract-level framing then proceed as if the posterior story were established ("The effect was also strongest in the posterior hippocampus"), with supporting citations to adult work, theoretical models and the rodent dorsal hippocampus. The anterior confidence interval (–0.064 to 0.431) overlaps most of the posterior one (0.139 to 0.629).

A voxel-wise GLM, labeled exploratory, shows bilateral posterior hippocampal voxels at P < 0.01 uncorrected across all infants and in older infants. The only cluster to survive correction (threshold-free cluster enhancement, P < 0.05) is right orbitofrontal cortex in older infants. Medial temporal lobe cortex outside the hippocampus, which carries item recognition in adults, shows nothing in the whole-brain map or in dedicated ROIs (fig. S10).

Uncorrected P < 0.01 across a whole brain produces false positive clusters. The authors do not rely on it; the hippocampal ROI analysis is the main result, and the whole-brain map is presented as a mirror of it, which is the right order. The orbitofrontal finding is interpreted through a study of children aged 4 to 8. The absence of MTL cortex is explained after the fact by task difficulty.

The paper's own wording is careful. The abstract says post-encoding mechanisms "may be more responsible." The discussion says the memories were fleeting, that the effect weakened with longer lags inside a 100-second window, that familiarity preferences signal partial encoding, and that "it remains possible that infantile amnesia may be partly attributable to real but impoverished encoding in infancy." The conclusion claims a boundary condition: accounts of infantile amnesia that assume a broad failure of hippocampal encoding are now harder to hold.

What was measured is a longer look at a photograph roughly one minute after a two-second exposure, in half the sample, on a within-infant basis. Nothing in the study tests whether any of these traces lasts a day, a week or into childhood. The retrieval account comes from mouse work in which infant engrams were reactivated optogenetically years of mouse-life later; this study supplies the human encoding half of that story and none of the persistence half. The gap between paper and packaging is the clearest bias in the record. Science's editor's summary says retrieval deficits "are likely responsible" for infantile amnesia. The Science press package says infantile amnesia is "more likely caused by memory retrieval failures." CNN's headline says the study "explains why" you cannot remember being a baby. The authors themselves told NPR that longer looking can also mean the infant is still learning the picture.

26 infants with usable data, demographics in table S1 (unread). Awake infant fMRI typically loses a large share of sessions to motion and fussing; the paper says only that excluded trials were dropped for movement or not looking, and that the number of usable encoding-test pairs was "already modest." The authors name attention, socioeconomic status and caregiver relationships as factors in infant memory they could not address, and call for larger samples.

Without the supplement, the funnel from infants recruited to infants analyzed is unknown here, and so is the number of trials each infant contributed. Both matter for a within-participant design where a handful of trials per infant are sorted by a noisy looking measure. A Yale Baby School sample is a New Haven convenience sample. None of this is hidden; it is in a file that was not readable from this desk.