Perturbing a temporarily unattended memory

Research date: 2026-09-22.

Citation and status. Nathan S. Rose, Joshua J. LaRocque, Adam C. Riggall, Olivia Gosseries, Michael J. Starrett, Emma E. Meyering, and Bradley R. Postle. Reactivation of latent working memories with transcranial magnetic stimulation. Science 354(6316), 1136–1139, 2 December 2016. DOI; PubMed; PMC. Peer-reviewed primary human experimental report; received 3 August and accepted 31 October 2016.

Sources and exact reading scope. Read the complete main text, all main figure captions, all supplementary Methods and Results, Figures S1–S6 and captions, Table S1, and Movie S1's caption in the author-hosted bundle. Its 27 pages contain four journal-layout pages followed by a 23-page textual supplement, including references. Visually inspected all main and supplementary figures and Table S1, and the pages containing sample counts and critical statistics. The journal-layout copy retains an embargo notice; exact byte identity with a current publisher download was not established. Also read the main prose of the institutional NIH manuscript, which has 12 pages and identifies itself as an author manuscript available through PMC from June 2017. Bibliographies were retained, not read as additional studies. Movie S1 itself was not retrieved or viewed. No data, code, or statistical reanalysis was performed. Publisher access failed; the Europe PMC API identifies this article as outside its open-access full-text service. Public author/institutional access does not establish an unrestricted reuse license.

Local main-plus-supplement PDF, extracted text, manuscript PDF, manuscript text, and provenance. The supplement's PDF text extraction substitutes punctuation for some spaces and hyphens; visual checks resolve the critical content.

Factual core, under 200 words. Participants retained two externally supplied items from different categories: faces, words, or motion directions. A cue prioritized one item for recognition while the other could still become relevant. Category decoding declined for the unattended item and returned when it was recued. During EEG recording, single-pulse TMS was followed by renewed evidence for the unattended category. A second experiment detected reactivation after the first cue but not after a later cue that made the uncued item irrelevant. A separate behavioral experiment reported more false alarms to unattended-item lures in a TMS group than in controls on the first probe. These results link task relevance, perturbation, category decodability, and recognition interference. They do not decode the individual face, word, or motion direction, measure synaptic weights, or establish recovery of unknown autobiographical information. Repeated sessions, small neural samples, broad temporal smoothing, between-group behavioral controls, and an unresolved sample-count discrepancy constrain stronger interpretations.

Design, sample, and analysis audit. Page numbers below refer to the 27-page author bundle.

What is decoded and held out. The fMRI localizer uses 72 trials, 24 per category, with regularized logistic classification and leave-one-trial-out validation. The main fMRI experiment applies delay-trained localizer classifiers to the second task. The supplement does not clearly document nesting the top-2,000-voxel selection inside each localizer validation fold; this is a reporting boundary, not proof of leakage into the separate task. Localizer searchlights choose stimulation regions for Experiments 2–3; fMRI-trained weights are not transferred to EEG.

EEG classifiers use 60 channels and 34 frequency features per channel. They are trained on multiple time points from other trials of the same task and tested on an entire withheld trial. Thus withheld time points from that trial are not knowingly mixed into training. However, z-scoring uses all trials and time points, including test features; this is not fully held-out prospective preprocessing. The stated 191-training-trial description fits Experiment 3's 192 trials but does not separately explain Experiment 2's 164-trial sessions. No participant-held-out or individual-item-held-out result is presented.

The target is whether a category was present, with AUC calculated separately when that category supplied the attended or unattended item versus when it was absent. AUC above .5 establishes relative discrimination under this design, not calibrated item probabilities or literal absence of any activation for the irrelevant category. Experimenters already know the presented items and cues; decoding does not demonstrate information unavailable from that task record. Equal categories and counterbalanced relevance make this useful physiology, but not a conditional extraction endpoint beyond all allowed inputs.

Timing and perturbation controls. In Experiment 2, TMS occurs 2.5 seconds after the first cue's offset, at a category-selective target held fixed within a session. Experiment 3 uses right precuneus and pulses after both cues, jittered 2–3 seconds after cue offset. Experiment 4 also uses precuneus, with fixed 2.5-second timing after both cues. The second cue occurs later in the trial, after a probe and response; relevance is therefore not the only difference between the two pulse periods.

Sound masking, artifact interpolation or median filtering, and ICA address TMS contamination. TMS occurs for all category conditions in the EEG experiments, and Experiment 3 fixes the target across categories; these controls reduce a simple category-specific pulse explanation. There is no matched sham/no-pulse EEG contrast reported. The general observation that every condition received TMS does not by itself eliminate all state-dependent sensory or residual-artifact explanations.

Crucially, spectral observations use 0.5-second windows and are averaged with two preceding and two following observations, yielding the authors' stated 2.5-second window. Adjacent plotted points are dependent, and apparent onset/offset cannot be interpreted with the original EEG sampling resolution. Tests compare post-pulse decoding with chance; a significant value after TMS is not automatically a significant increase from the preceding value. Targeted versus nontargeted Bayes factors likewise do not substitute for their direct difference test. No time-series multiplicity correction is specified in the read Methods. The band analysis associates the reactivation result with 16–24 Hz beta features; it does not selectively intervene on beta activity to prove that frequency band is necessary.

Behavior and mechanism. In Experiment 4 the reported first-probe unattended-lure correct-rejection rates are .86 with TMS and .92 in controls; second-probe rates are .85 and .87. The first comparison reports t(39)=2.33, p=.01, with a Bayes factor near 3.5; the second reports t(39)=.48, p=.32. The supplement reports a stimulus-type-by-TMS interaction, but no probe-number-by-TMS interaction establishing that the two TMS contrasts differ. The textual description implies a 50/30/20 split among attended matches, attended-category nonmatches, and unattended-item lures; its phrase “of the 50%” is arithmetically awkward and should not be interpreted as additional conditional percentages without checking the task code.

The comparison is a behavioral interference result, not memory enhancement. The Methods do not specify random allocation or sham matching between the TMS and behavioral-only groups. Because the unattended item belongs to a different category from the attended item, exact unattended lures are not compared with matched novel lures from the unattended category. The behavioral result therefore supports increased intrusion but does not uniquely resolve the authors' acknowledged category-context alternative to item reactivation. The neural and behavioral effects also come from different experiments, not a demonstrated within-trial mediation pathway.

The data are consistent with a latent, relevance-sensitive state that can affect a later response. Nondecodability by these category classifiers does not establish zero neural activity or loss of all other decodable codes. Conversely, failure to reactivate a dropped item does not establish erasure. Short-term synaptic plasticity is a proposed substrate, not directly measured here; changed attention, representation, accessibility, and weak or differently coded ongoing activity remain distinct mechanistic possibilities. Supplementary source-localized current and complexity measures did not distinguish the memory states, and the authors leave precise anatomical storage hypotheses unresolved.

Extraction and transfer implications — our reasoning. A readout failure can reflect a mismatch between measurement and state rather than missing information. A perturbation can change both the channel and subsequent behavior, so its value must be assessed jointly with induced errors and retention, rather than treating recovered classifier evidence as a passive snapshot. This study is neither a low-burden EEG-only extraction demonstration nor evidence that TMS strengthens a person's autobiographical memories.

A future informational test would require fresh within-category item details, stimulus/cue-matched baselines, item-held-out validation, training-only preprocessing, and participant-level uncertainty. A perturbation comparison would additionally need matched sham or other appropriate controls, a timing analysis that cannot use future samples, and separate tests of exact-item confusion versus category context. Report immediate category evidence, later item accuracy, source confusion, and retained information as different endpoints. A useful artificial analogue is to test stored records that become inaccessible under one query operation and accessible under another; success would test that implementation's access policy, not prove synaptic storage or biological fidelity.