Neural Traces of Forgotten Memories Persist in Humans and are Behaviorally Relevant

Reviewer #1 (Public review):

Summary:

This manuscript presents an ambitious attempt to examine whether episodic memory traces ("engrams") of forgotten associations persist in the human brain and whether these traces continue to influence behavior implicitly. Using 7T fMRI, the authors track 96 one-shot face-object associations across learning, 30-minute retrieval, and 24-hour retrieval, complemented by a recognition test. Participants classify each memory as sure, unsure, or guess, enabling an operational dissociation between consciously accessible and inaccessible memories.

Strengths:

The study addresses a timely and theoretically important question arising from rodent engram research, i.e., whether forgotten human memories leave detectable neural signatures. The use of high-resolution 7T fMRI, representational similarity analysis (RSA), and gPPI connectivity analyses aims at a detailed systems-level perspective. The results suggest that correct guess responses (i.e., when participants believe they are guessing) are accompanied by hippocampal activity and connectivity patterns that correlate with behavioral performance, potentially pointing to residual memory traces. The study also presents evidence for divergent consolidation trajectories: consciously accessible memories become more neocortically distributed after sleep, whereas inaccessible memories exhibit strengthened hippocampal signatures.

Weaknesses:

Despite the methodological rigor, some interpretational issues merit caution. First, the reliance on participants' subjective "guess" reports to categorize trials as forgotten is problematic. Guess responses at the 30-minute retrieval were at chance level, whereas guess responses during recognition were above chance; interpreting both as "implicit episodic memory" may conflate different mechanisms (episodic retrieval, familiarity, associative priming).

Second, several analyses raise concerns about circularity or insufficient independence, for example, when contrasting correct vs. incorrect guess trials to locate "engram" activity and then correlating that activity with guessing accuracy. Similarly, the behavioral analyses are fragmented (multiple t-tests across conditions) rather than using a factorial model that accounts for dependencies among confidence levels and timepoints.

Third, the choice to include only "sure" and "guess" responses discards a substantial portion of trials ("unsure"), reducing power and complicating interpretation, especially given that unsure responses show above-chance performance.

Finally, the study's two-scanner-sequence design (small-FOV vs. whole-brain) is challenging as it complicates comparisons across analyses, especially when some critical results (e.g., hippocampal reinstatement patterns) do not consistently replicate across sequences.

Conclusion:

Overall, the manuscript provides preliminary evidence that neural traces of forgotten episodic memories might persist in humans and could guide behavior in the absence of conscious awareness. While interpretational caution is warranted, especially regarding the nature of "guess"-based retrieval and the independence of neural contrasts, the study makes a valuable contribution to debates on engram persistence, systems consolidation, and the role of consciousness in episodic memory.

Reviewer #2 (Public review):

Summary:

The goal of the experiment was to identify the fMRI neural correlates of persistence and recovery of forgotten memories. A forgotten memory was defined behaviorally as successful learning, followed by failure in a recall format task, followed by next-day success in a recognition format task. The comparison is to memories that were not forgotten at any stage of the task. Various univariate, connectivity, and multivariate analyses were used to identify neural correlates of forgotten memories that were recovered, that remained forgotten, and successful memory. Some claims are made about how activity of the "episodic memory network" predicts the persistence of forgotten memories.

Strengths:

Studies on the persistence of forgotten memories in rodent models have been used to make some novel claims about the potential properties of engrams. Attempting similar research in humans is a laudable goal.

Patterns of behavioral responses are consistent across subjects.

Weaknesses:

I do not find that the fMRI results fit the narrative provided.

A major issue is that primary results do not replicate across the two fMRI datasets that were collected using the same task. For example, hippocampal activity associated with correct responses (confident and guess) was identified in the group receiving the fMRI scan that used a small FOV, but not in the group that received an fMRI scan of the whole brain, for both 30-min and 24-hr delays (lines 202-217). This suggests that the main findings are not even replicable internally within the same experiment. There is no reasonable justification for this.

Next, most of the reported fMRI findings do not meet reasonable thresholds for statistical significance. In many places, the authors acknowledge this in the text by saying that a difference in the fMRI metric "tended towards significant correlation" or that comparisons "revealed non-significant mean value comparisons". It is not clear why these non-significant findings are interpreted as though they are positive findings. Beyond that, many of the reported findings are not meeting the threshold (i.e., p=0.058), without any acknowledgement that they are marginal. Beyond that, the majority of comparisons that are interpreted in the main text are not significant based on the companion information provided in the supplementary tables. That is, they are totally non-significant when using FWE or FDR correction at either the cluster or peak levels.

Beyond this, the supplementary tables indicate that "clusters identified solely within white matter regions have been excluded." The fact that there are any findings in white matter to ignore indicates that the statistical thresholds are inappropriate. It's tantamount to seeing activation in the brain of a dead fish.

The overall picture based on these factors is that the statistical tests did not use sufficiently stringent safeguards against false positives given the multiple comparison problem that plagues fMRI. So, there are tons of false positives, which are being selectively interpreted to tell a particular story. That is, each comparison yields lots of findings in many brain area, and those that do not fit the particular narrative are being ignored (including those in white matter). What's more, when the small FOV fMRI scan is done, the imaging volume is centered on the hippocampus and its close network, so all false positives appear to be exactly in those brain regions about which the authors want to make conclusions. When throwing darts, you will always hit a bullseye if that is all that exists. The fact that the same comparisons done in the companion whole-brain dataset do not yield the same results is telling: the analysis plan is not sufficiently rigorous to yield findings that are replicable.

Further, I think that it is highly debatable whether the task measures the recovery of forgotten memories at all. Forgotten memories are defined as those that fail when tested using a recollection format but succeed when tested using a recognition format. The well-characterized distinction between recollection and recognition is thus being construed as telling us something about the fate of engrams. I think the much more likely alternative is that "forgotten" memories are just relatively weak memories that don't meet whatever criteria subjects typically use when making recollection judgments, and not some special category of memory. In terms of brain activation, they seem for the most part to follow the pattern of stronger memory, but weaker.

Finally, many hypotheses are used as though they are proven. For instance, fMRI activity patterns are called "engrams" even though there are no tests to determine whether they meet reasonable criteria that have been adopted in the engram literature (e.g., necessity, sufficiency). Whatever happens over the 24-hour delay is called "consolidation" even if there is no test that consolidation has occurred. Etc. It becomes hard to differentiate what is an assumption, versus a hypothesis, versus an inference/conclusion.