My son was fourteen months old and I was on the living room floor rebuilding the same three-block tower for what felt like the fortieth time that week. Someone said the thing everyone says. He won't remember any of this.
They meant it kindly. It is also the most load-bearing belief parents of zero to two year olds hold, and it licenses a lot of coasting. If none of it registers, the first two years are logistics. Feed, sleep, keep alive, start paying attention around three.
In March 2025, a team at Yale published direct neural evidence that the belief is aimed at the wrong part of the process.
What the scan actually showed
The paper is Yates and colleagues, "Hippocampal encoding of memories in human infants," in Science, out of Nicholas Turk-Browne's lab at Yale. Twenty-six infants produced usable awake fMRI data, aged 4.2 to 24.9 months, half under one year and half over.
The design is simple. Show an infant a photograph they have never seen. About a minute later, show it again beside a new one from the same category. The scanner records what the hippocampus did the first time. The readout is whether the infant then looks longer at the one they already saw.
The result is trial by trial, not child by child. Stronger hippocampal activity while an image was first viewed predicted that the same infant would later look longer at that image. Encoding, caught in the act.
Age split it cleanly. In the older group, 12 to 24 months, the effect was strong and significant (P less than 0.001). In the younger group, 4 to 9 months, there was nothing (P = 0.898), and the difference between the groups held up (P = 0.002). The authors' own line: "episodic encoding in the current study was observed only in infants older than 12 months."
Then the part that matters, from the abstract. The availability of encoding mechanisms during a period of life "that is later lost from our autobiographical record implies that postencoding mechanisms, whereby memories from infancy become inaccessible for retrieval, may be more responsible for infantile amnesia."
Translated: something got recorded. What is most likely broken later is playback.
The caveats, stated plainly
Twenty-six infants. Awake infant fMRI is hard enough that small samples are the norm, but it is still 26.
The gap between seeing an image and being tested on it averaged 59.9 seconds, with three to twelve other items in between. Long enough, the authors argue, to be past working memory. Nowhere near long enough to say anything about a memory lasting a week, a year, or a lifetime.
Each image was shown once. The effect was clearer at shorter lags than longer ones, and the authors write that "the fleeting nature of these memories may be expected given the brief, one-shot exposure during encoding." They also found no overall preference for familiar images across the sample, which is a fair reminder of how noisy looking time is as a measure.
So nobody has shown that a specific memory from your child's first year survives into adulthood. Turk-Browne calls that "the radical, almost sci-fi possibility" his lab is only beginning to entertain.
The mechanism evidence is animal work. Guskjolen and colleagues, Current Biology, 2018: mice trained at 17 days old showed near-zero freezing when returned to the same context 15 to 90 days later, apparently having forgotten. Optogenetically stimulating the dentate gyrus neurons tagged during that learning brought the behavior back, up to three months after training. Storage intact. Retrieval not.
The honest version is not "your baby remembers her first birthday." It is that adult amnesia for infancy looks less like an empty tape and more like a lost index.
Encoding starts well before that
The 12-month threshold applies to episodic memory, the memory for particular events. A different system is running much earlier.
From the same Yale lab, Ellis and colleagues in Current Biology, 2021, scanned 17 infants aged three months to two years. The hippocampus was more active when a sequence of images had an order to be learned than when the order was random, and that activity had no relationship to age, which is why the authors put the onset at the youngest infants they had, around three months.
The behavioral version has been on the record since 1996. Saffran, Aslin and Newport, in Science, played 8-month-olds two minutes of a continuous stream of synthesized syllables. No pauses, no stress cues, no instruction, no reward. The only information available was the odds of one syllable following another. Two minutes was enough for the infants to segment recurring "words" out of the stream, based, in the authors' words, "solely on the statistical relationships between neighboring speech sounds."
That is the system that is on from the beginning. It does not store events. It stores structure.
The strangest evidence comes from people who cannot speak a word of it
Pierce and colleagues, PNAS, 2014, scanned 21 people who had been adopted from China into French-speaking families in Quebec. Mean age at adoption: 12.8 months. Roughly 12.6 years since any exposure to Chinese. They spoke only French and had no conscious recollection of the language.
Given a Chinese lexical tone task, their brains patterned with Chinese-French bilinguals who had heard Chinese continuously since birth, activating mainly the left hemisphere. French monolinguals who had never heard Chinese activated the right hemisphere only.
Twelve months of hearing a language, then twelve years of silence, and the signature is still there. Not as anything anyone can report. As a shape the system got built into.
What this changes on a Tuesday
If what persists is structural rather than episodic, the practical implication runs opposite to instinct.
The instinct, and it is the one social media trains hardest, is to buy memorable. The zoo day. The trip. The one big thing you will have photos of.
Statistical learning does not run on peak intensity. It runs on frequency. What is getting encoded is not the aquarium. It is the repetitions: you narrating the walk to the car, the same book at the same hour, the same turn-taking rhythm where you say something and wait and they answer with a sound and you answer back. That is the part that feels least worth doing while you are doing it, because there is no visible feedback.
Three things follow.
Talk into the apparent void. No response is not evidence of no effect. A gap between what the brain is doing and what the behavior shows is exactly what these studies are built to detect.
Prefer the repeated ordinary over the singular special when you have to choose, and you usually do have to choose, because attention is the scarce input in a day with a toddler in it.
Stop grading yourself on whether it landed. You will not get a receipt. Zero to two is the one stretch of parenting with no scoreboard, which is precisely why "she won't remember any of this" survives as an argument. It is unfalsifiable from the couch. It is also not what the imaging says.
One boundary worth naming. None of this describes any individual child, and none of it is a way to evaluate your own. Anything you are actually wondering about with your kid is a conversation with your pediatrician, who can see the child in front of them. This is about what the research says about the window.
What we do with it
The reason I keep coming back to this study is not the headline. It is the reframe. The first two years are not a waiting room before the real work starts. They are the period with the highest ratio of input to visible output, which makes them the easiest to under-invest in.
Prodigy is built on that assumption. Not that you should do more, but that the ordinary repeated day is the input that counts, and that it is hard to sustain something you get no feedback on. So Prodigy works at that grain: what to do this week, with the child you have, at the age they are, and a record of what you actually did. The window gives you no receipt. The least we can do is keep the log.