Fast network oscillations during non-REM sleep support memory consolidation

The neocortex is disconnected from the outside world during sleep, which has been hypothesized to be relevant for synaptic reorganization involved in memory consolidation. Fast network oscillations, such as hippocampal sharp-wave ripples, cortical ripples, and amygdalar high-frequency oscillations, are prominent during non-REM sleep. Although these oscillations are thought to be generated by local circuit mechanisms, their occurrence rates and amplitudes are modulated by thalamocortical spindles and neocortical slow oscillations during non-REM sleep, suggesting that fast network oscillations and slower oscillations cooperatively work to facilitate memory consolidation. This review discusses the recent progress in understanding the generation, coordination, and functional roles of fast network oscillations. Further, it outlines how fast network oscillations in distinct brain regions synergistically support memory consolidation and retrieval by hosting cross-regional coactivation of memory-related neuronal ensembles.


Introduction
Hippocampal sharp-wave ripples (SWRs; 140-230-Hz transient oscillations coupled with irregularly occurring sharp waves) are selforganized patterns that emerge from the extensive excitatory recurrent system of the CA3 and represent synchronous population activity in the CA3-CA1-subicular complex and entorhinal cortex (Buzsaki, 2015). Sequential activity that occurs while awake is replayed, supported by SWRs, in a time-compressed manner during non-REM sleep (Buzsaki, 2015). It has been proposed that SWRs mediate the transfer of newly acquired hippocampal information to brain-wide circuits during waking and non-REM sleep, thereby supporting memory consolidation (Buzsaki, 2015).
Accumulating evidence indicates that fast network oscillations in the frequency range of hippocampal ripples are prominent during non-REM sleep in brain areas outside the hippocampal formation, such as the neocortex (Grenier et al., 2001;Khodagholy et al., 2017;McKenzie et al., 2020), amygdala (Ponomarenko et al., 2003), and lateral septum (Tingley and Buzsaki, 2020). This review discusses recent progress in understanding the generation, coordination, and functional roles of hippocampal SWRs, cortical ripples, and amygdalar high-frequency oscillations (HFOs). Further, we outline how these fast network oscillations support cross-regional coactivations of memory-related neuronal ensembles and memory consolidation.

Theta and non-theta states
Local field potential (LFP) patterns along the laminar structure of the hippocampus, with projections from different brain regions terminating in different layers, reflect a variety of behavior-dependent network dynamics (Buzsáki, 2006). Two brain states associated with distinct LFP oscillations and population synchrony, which show an antagonistic relationship with each other, are prominent in the hippocampus (Buzsaki, 2002(Buzsaki, , 2015Mizuseki and Buzsaki, 2014;Miyawaki, 2017, 2019). During preparatory behaviors (e.g., ambulation, exploration, rearing, and sniffing) and REM sleep, continuous theta oscillations (6-10 Hz) dominate, and SWRs are rarely observed in the hippocampus. The brain state associated with theta oscillations is called the theta state. In contrast, during immobility, consummatory behaviors (e.g., eating, drinking, and grooming), and non-REM sleep, irregularly occurring SWRs are prominent, and theta oscillations are suppressed in the hippocampal CA1 area. The associated brain state is classified as the non-theta state.
Theta and non-theta states show distinct neuromodulatory tones. Notably, the acetylcholine level in the hippocampus is higher during REM than during non-REM sleep (Buzsaki, 2015). Systemic application of the anticholinergic drug atropine increases the amplitude of SWRs (Buzsaki et al., 1986), and optogenetic stimulation of cholinergic neurons in the medial septum suppresses the occurrence of hippocampal SWRs (Vandecasteele et al., 2014).

Mechanisms generating hippocampal SWRs
An SWR is a high-frequency (140-230 Hz) ripple oscillation that occurs transiently in the CA1 pyramidal cell layer and is accompanied by a negative deflection called a sharp wave in the CA1 stratum radiatum (Fig. 1), which is generated by the current sink that reflects synchronized glutamatergic input from CA3 pyramidal neurons to the apical dendrites of CA1 pyramidal neurons (Buzsaki, 2015). The firing rate of CA3 pyramidal neurons increases just before SWRs (Buzsaki, 2015), and optogenetic silencing of CA3 pyramidal neurons strongly reduces the incidence of CA1 ripples (Davoudi and Foster, 2019;Yamamoto and Tonegawa, 2017), suggesting that CA3 is the main CA1 ripple driver. However, the frequency of CA3 ripples is slower than that of CA1 ripples, and CA3 and CA1 do not show robust ripple oscillatory phase coherence, suggesting that ripples in the CA1 are not inherited from CA3 in a cycle-by-cycle manner but generated in the local circuit of CA1 (Buzsaki, 2015;Sullivan et al., 2011). Furthermore, ripples, though with a significantly lower frequency, are observed in CA1 when the CA3 input is chronically silenced (Middleton and McHugh, 2016;Nakashiba et al., 2009), and subsets of ripples are not coupled with a sharp wave (Ramirez-Villegas et al., 2015). Thus, ripple generation in CA1 may be controlled by mechanisms involving structures other than CA3, such as synchronous depolarizing input from CA2 (Alexander et al., 2018;Boehringer et al., 2017;He et al., 2021;Oliva et al., 2016Oliva et al., , 2020 or the entorhinal cortex (Yamamoto and Tonegawa, 2017). Any strong depolarizing input to the local circuit can trigger a ripple response regardless of its origin. In line with this notion, non-rhythmic optogenetic depolarization of CA1 pyramidal cells induces fast network oscillations, suggesting that the intrinsic circuitry of CA1 is sufficient for ripple generation .
The exquisite balance between excitation and inhibition is maintained by the interplay between hippocampal pyramidal cells and interneurons during SWRs (Buzsaki, 2015;Ylinen et al., 1995). An in vivo intracellular study showed that hyperpolarizing inhibition counterbalanced the dendritic excitation of pyramidal neurons during SWRs (English et al., 2014). Moreover, inhibitory post-synaptic currents (IPSCs), but not excitatory post-synaptic currents, are phase-locked to the LFP ripples and correlate with their amplitude, indicating that inhibition is the primary determinant of ripple frequency (Gan et al., 2017). In vivo local picrotoxin (a GABA-A receptor antagonist) administration was shown to inhibit ripple oscillations, indicating that GABA-A receptor-mediated inhibition is required for ripple generation . The firing of many types of hippocampal inhibitory neurons is positively or negatively modulated by SWRs (Freund and Buzsaki, 1996;Klausberger and Somogyi, 2008;Szabo et al., 2022). Among them, CA1 parvalbumin (PV)-positive, fast-spiking GABAergic interneurons are the best candidates for the source of inhibition-controlling ripples because they show strong phase-locking to ripples and can pace CA1 pyramidal cell spiking (Buzsaki, 2015;Stark et al., 2014). Optogenetic inhibition of PV-positive interneurons reduces the ripple duration and disrupts the IPSC phase-locking to the ripple (Gan et al., 2017), suggesting that the GABA-A receptor-mediated current activated by PV-positive interneurons controls the generation and frequency of ripple oscillations in the CA1. Tight ripple-phase coherence among distant CA1 regions (Patel et al., 2013) is also mediated, at least partly, by GABAergic interneurons. When fast network oscillations are optogenetically induced in the CA1 by non-rhythmic optical stimulation, the firing of distant interneurons, but not pyramidal neurons, is synchronized with local fast network oscillations . Moreover, ripple-band coherence between nearby electrodes is disrupted by local injection of picrotoxin, indicating the potential role of GABA-A receptor-mediated mechanisms in coherent ripple activity in anatomically distant locations . In hippocampal slice experiments, stimulation of PV-positive basket cells induces ripples even with excitation blocked (Schlingloff et al., 2014), further supporting the notion that local PV-positive interneurons play a pivotal role in ripple oscillations. In addition to generating and coordinating ripple oscillations, inhibition selects pyramidal cells participating in SWRs and precisely controls their spike timing during SWRs (Noguchi et al., 2022;Valero et al., 2017Valero et al., , 2015.

Brain-wide impact of hippocampal SWRs
Whole-brain blood-oxygen-level-dependent functional magnetic resonance imaging combined with electrophysiology revealed that many cortical areas are more active during hippocampal SWRs (Logothetis et al., 2012). On the other hand, the subcortical regions responsible for sensory information processing, such as the thalamus, are silent during hippocampal SWRs, minimizing interference with offline information transfer from the hippocampus to the neocortex (Logothetis et al., 2012;Yang et al., 2019). Moreover, electrophysiological recordings with high temporal and spatial resolution showed that the firing rates of many neocortical and thalamic regions are enhanced and suppressed, respectively, during hippocampal SWRs (Nitzan et al., 2022). The subiculum, a major output structure of the hippocampal formation, receives direct input from the hippocampus CA1 and projects to various cortical and subcortical regions (Matsumoto et al., 2019;Mizuseki and Kitanishi, 2022). Consistently, the thalamus-projecting subicular neurons are selectively suppressed, whereas the majority of subicular neurons projecting to other regions, such as the retrosplenial cortex (RSC), are activated during SWRs, indicating that SWRs control information distribution from the hippocampal formation to the downstream areas in a target-specific manner (Kitanishi et al., 2021). Furthermore, SWRs in the dorsal and ventral hippocampus occur largely asynchronously during non-REM sleep (Patel et al., 2013) and even more asynchronously during awake states and have distinct effects on the neuronal activity in target networks (Sosa et al., 2020). In addition to the septo-temporal axis of the origin of SWRs, the effect of SWRs on the activity patterns of extrahippocampal structures depends on the amplitude and synchrony of SWRs (Nitzan et al., 2022). Wide-field optical imaging of voltage and glutamate activity of the neocortex showed that the neocortex tended to activate sequentially from medial to lateral regions around SWRs on a time scale of approximately 30 ms (Karimi Abadchi et al., 2020). Both local inhibition in the RSC and excitatory input from the thalamus to the RSC are suppressed during SWRs (Chambers et al., 2022). On the other hand, the RSC, which receives direct SWR-associated excitatory drive from the subiculum (Kitanishi et al., 2021;Nitzan et al., 2020), was most strongly activated during hippocampal SWRs, suggesting that the RSC promotes broadcasting of hippocampal representations to the rest of the neocortex (Karimi Abadchi et al., 2020). Nevertheless, the increase and decrease in neuronal activity in target areas during/after SWRs should not be equated with excitation or inhibition of the target areas by the SWRs because of additional, simultaneously occurring brain state changes (Nitzan et al., 2022).

Hippocampal SWRs mediating memory consolidation
The "two-stage model" proposes two distinct stages of memory formation: the "online" rapid memory acquisition that occurs in the theta state during wakefulness and the "offline" memory consolidation that is promoted through reactivation of acquired memory traces in the nontheta state during subsequent rest and sleep, when the brain is disengaged from the external environment (Buzsaki, 1989(Buzsaki, , 2015. According to this hypothesis, in the theta state during wakefulness, input from the neocortex to the hippocampus induces transient synaptic reorganization in the hippocampus, where newly acquired information is temporally retained as a labile form of the memory trace. During the subsequent non-theta state in rest and sleep, hippocampal highly synchronized population bursts associated with SWRs facilitate the transfer of this temporally retained information from the hippocampus to the neocortex and induce long-term synaptic reorganization in the circuit, thereby transforming labile memory traces into enduring ones (Buzsaki, 1989(Buzsaki, , 2015. In support of the hypothesized role of SWRs in memory consolidation, offline neuronal reactivation during SWRs has been shown to induce synaptic reorganization (Buzsaki, 2015;Norimoto et al., 2018;Sadowski et al., 2016) and a net decrease of hippocampal firing across sleep (Grosmark et al., 2012;Miyawaki and Diba, 2016). Moreover, the incidence of SWR events during non-REM sleep has been shown to increase after odor-reward association and spatial memory tasks and positively correlate with subsequent performance improvement (Eschenko et al., 2008;Ramadan et al., 2009). Furthermore, hippocampal neuronal ensembles associated with novel or rewarded experiences reactivate more strongly than those associated with familiar or unrewarding experiences (McNamara et al., 2014;Mizunuma et al., 2014;O'Neill et al., 2008;Singer and Frank, 2009). Replays during sleep following exploration of novel environments persist longer than those of familiar environments (Giri et al., 2019), underscoring the cardinal role of SWRs in learning and memory.

Temporal coordination of hippocampal SWRs and slower oscillations outside the hippocampus
Neocortical slow oscillations (<1 Hz), reflecting synchronized fluctuations of the membrane potential of neocortical neurons across layers and neocortical regions, are hallmark activity patterns observed in anesthetized and sleeping animals (Steriade, 2003). During slow oscillations, the membrane potential of virtually all neocortical neurons synchronously alternate between a hyperpolarized DOWN state and a depolarized UP state (Steriade, 2003). Non-REM sleep is characterized by neocortical delta waves (0.5-4 Hz) (Sirota and Buzsaki, 2005;Steriade, 2003) and thalamocortical sleep spindles (9-18 Hz waxing-and-waning waves (De Gennaro and Ferrara, 2003;Fernandez and Luthi, 2020;Sirota and Buzsaki, 2005;Steriade, 2001;Steriade, 2003;Sullivan et al., 2014)). Neocortical delta waves are associated with transient (150-500 ms) and widespread cessation of spiking activity of both principal neurons and interneurons in all cortical layers and followed by epochs (0.3 − 1 s) of sustained activity. Thus, the delta waves and slow oscillations are not separate network patterns, but the delta wave can be regarded as the DOWN portion of the slow oscillation (Sirota and Buzsaki, 2005) (but see (Kim et al., 2019), which suggests that cortical slow oscillations and delta waves are distinct network patterns and have distinct roles in memory).
During non-REM sleep, hippocampal SWRs and neocortical oscillations in slower frequency bands are temporally coordinated, supporting the hypothesized role of SWRs in information transfer and memory consolidation (Buzsaki, 2015;Miyawaki, 2017, 2019). Thalamocortical sleep spindles modulate hippocampal SWRs; both events are modulated by neocortical slow oscillations; and all three oscillations are modulated by the phase of the ultra-slow rhythm ((Aladjalova, 1957); approximately 0.1 Hz) Sirota et al., 2003). Such cross-frequency coupling of rhythms, in which lower-frequency oscillations modulate higher-frequency oscillations, constitutes the computational basis of the brain's hierarchical organization of multiple timescales (Buzsáki, 2006;Sirota et al., 2003). Hippocampal SWRs tend to occur during the transition from DOWN-to-UP states and UP-to-DOWN states (Battaglia et al., 2004;Isomura et al., 2006;Karimi Abadchi et al., 2020;Miyawaki and Mizuseki, 2022;Molle et al., 2009;Peyrache et al., 2011Peyrache et al., , 2009Sirota et al., 2003) of slow oscillations (Steriade et al., 1993a; and coincide with thalamocortical sleep spindles (Clemens et al., 2007(Clemens et al., , 2011Molle et al., 2009;Ngo et al., 2020;Siapas and Wilson, 1998;Sirota et al., 2003). The incidence of long-duration SWRs has been shown to increase in situations demanding memory, and optogenetic induction of prolonged SWRs recruits additional neurons to SWR-related neuronal sequence and improves memory (Fernandez-Ruiz et al., 2019). Importantly, recordings from patients with pre-surgical epilepsy revealed that concurrent spindle power increments of the hippocampus and neocortex time-locked to hippocampal SWRs are pronounced during long-duration SWRs (Ngo et al., 2020). Furthermore, bundled hippocampal SWRs, which support replay of prolonged experience (Davidson et al., 2009), tend to occur during more extended neocortical UP states than isolated SWRs, suggesting that a longer temporal window is available for bundled SWRs for hippocampal-neocortical communications (Karimi Abadchi et al., 2020). In summary, temporal coordination between hippocampal SWRs and slower oscillations outside the hippocampus is suggested to be critical in memory consolidation.

Coordinated reactivations between the hippocampus and its partner regions supported by hippocampal SWRs
Hippocampal SWRs have been hypothesized to coordinate reactivations between the hippocampus and its partner regions (Buzsaki, 2015). The reactivation of neuronal ensemble firing patterns in the neocortex coincides with hippocampal SWRs on a coarse time scale (Jadhav et al., 2016;Ji and Wilson, 2006;Peyrache et al., 2009). In humans, visual stimulus-specific neocortical gamma activity observed during the encoding phase of visual stimulation is replayed during non-REM sleep after the encoding. This replay is time-locked to the hippocampal SWRs and correlated with remembered and forgotten items, indicating a prominent role of SWR-triggered cortical replay in memory consolidation (Zhang et al., 2018).
The prefrontal cortex receives direct input from the ventral hippocampus (Tovote et al., 2015). Cell assemblies formed in the medial prefrontal cortex during arousal are reactivated in a temporally compressed manner during non-REM sleep (Euston et al., 2007;Peyrache et al., 2009). Reactivation of the prefrontal cortical ensembles is positively correlated with the density of state transitions from DOWN-to-UP states (Johnson et al., 2010) and occurs most often near the beginning and end of the UP state when hippocampal SWRs are prominent (Peyrache et al., 2009). Moreover, as rats learn behavioral rules, neurons in the prefrontal cortex are recruited to cell ensembles coordinated with the hippocampal theta oscillations, and the same ensembles are preferentially reactivated during hippocampal SWRs in subsequent sleep (Benchenane et al., 2010).
The basolateral nucleus of the amygdala (BLA) is reciprocally connected with the ventral hippocampus and is important for fear memory (Tovote et al., 2015). A previous study found that coordinated reactivations of neuron pairs between the dorsal hippocampus and BLA that are activated during trials passing a corridor near the air puff are observed during non-REM sleep after the location-threat association task (Girardeau et al., 2017). These co-reactivations peak during hippocampal SWRs in non-REM sleep and are reinstated during test trials without the air puff. A subgroup of BLA neurons active during hippocampal SWRs preferentially participate in coordinated reactivation. These observations suggest that hippocampus-BLA coordinated reactivations hosted by hippocampal SWRs support consolidating emotional memory (Girardeau et al., 2017). Thus, hippocampal SWR-assisted coupled reactivation of the ensembles between the hippocampus and its partner regions may play a pivotal role in memory consolidation Miyawaki, 2017, 2019).
Recent "loss-of-function" and "gain-of-function" studies have revealed the importance of oscillation-related activities during sleep in memory consolidation. Disturbance of SWR-related hippocampal neural activity during sleep by electrical stimulation after spatial learning causes memory impairment (Ego-Stengel and Wilson, 2010;Girardeau et al., 2009). Moreover, offline reactivation of newly formed hippocampal cell ensembles, unlike that of previously developed ones, is positively correlated with future context-dependent reinstatement of these ensembles (van de Ven et al., 2016). Closed-loop optogenetic disruption experiments showed that ensemble reactivation during SWRs is required only when cell assemblies are newly formed and is progressively strengthened during the first exposure to a novel place (van de Ven et al., 2016). Furthermore, induction of delta waves and spindles in the neocortex by SWR-triggered electrical stimulation during post-training non-REM sleep enhances the coordination of hippocampal and prefrontal cortical activity and improves memory performance in a novel object recognition task, suggesting that the hippocampus-prefrontal cortex coupling around SWRs mediates memory consolidation (Maingret et al., 2016). In addition, optogenetic induction of spindles during the UP state of slow oscillations results in increased triple coupling of slow oscillation-spindle-ripple events and promotes consolidation of hippocampus-dependent memory during sleep, underscoring the critical role of the coordinated activity of slow oscillations, spindles, and hippocampal SWRs in memory consolidation (Latchoumane et al., 2017).
Similar to hippocampal ripples, ripple oscillations in the parahippocampal regions are also associated with sharp waves Buzsaki, 1994, 1996). In the neocortex, cortical ripples tend to be associated with the trough of local spindles, during which cortical neurons display depolarization (Averkin et al., 2016). Artificial optogenetic activation of CaMK2-positive cells in layer 5 of the somatosensory cortex  and VGluT2-positive bursting neurons in the subiculum  locally induces fast oscillations whose frequency is in the ripple band, suggesting that strong input to excitatory cells triggers the local network interaction that gives rise to fast oscillations (Buzsaki, 2015). The firing of PV-positive fast-spiking interneurons recruited during spindle-nested cortical ripples and phase-locked to the ripple oscillations follows the firing of pyramidal neurons with latencies of approximately 1-3 ms (Averkin et al., 2016), suggesting that the dynamic interplay between pyramidal neurons and interneurons plays a significant role in generating cortical ripples (Averkin et al., 2016;Buzsaki, 2015;Khodagholy et al., 2017;Nitzan et al., 2020).
Ripple oscillations, which occur concurrently with hippocampal SWRs, are prominent in retrohippocampal regions, such as the subicular complex (Chrobak and Buzsaki, 1996), medial entorhinal cortex (MEC) Buzsaki, 1994, 1996;Mizuseki et al., 2009), and RSC Opalka et al., 2020). Ripple oscillations along the CA1-subicular-entorhinal axes decrease in amplitude, decrease in phase coherence with CA1 ripples, slow in frequency, and process fewer cycles as a function of increasing distance from CA1 (Buzsaki, 2015;Chrobak and Buzsaki, 1996), suggesting that cycle-by-cycle ripples propagation decays as the distance from CA1 increases. Notably, MEC ripples are associated with a negative-going sharp wave in deep layers of MEC, reflecting depolarizing input by hippocampal SWRs-associated CA1 population burst Buzsaki, 1994, 1996). Similarly, RSC ripple is coupled with a local shape wave, which is temporally associated with hippocampal SWRs . Superficial, but not deep, RSC neuron firing is phase-locked to hippocampal ripple cycles . Optogenetic activation of subicular VGluT2-positive bursting neurons, which densely project to the layers 2/3 of RSC, induces ripple-band activity in the superficial but not deep RSC . Moreover, optogenetic silencing of subicular VGluT2-positive bursting cells terminals in RSC during hippocampal SWRs results in the reduction of ripple oscillations in the RSC , indicating that VGluT2-positive subicular bursting neurons mediate SWR-related activity from the CA1 to the RSC. Thus, hippocampal-subicular-entorhinal and hippocampal-subicular-RSC paths can broadcast hippocampal content associated with SWRs.

Coordination of hippocampal and cortical ripples supporting memory
A previous study has demonstrated that, during non-REM sleep, cortical ripples are prominent in association cortices, such as the medial prefrontal cortex, anterior cingulate cortex, RSC, and posterior parietal cortex (Khodagholy et al., 2017), which are reciprocally connected with the medial temporal lobe (Buckner et al., 2008;Burwell and Amaral, 1998), in comparison with the primary somatosensory, visual, and motor cortices. During non-REM sleep, hippocampal and cortical ripples often show a close temporal association (Khodagholy et al., 2017). As mentioned above, similar to hippocampal ripples, cortical ripples are temporally correlated with spindles and delta waves. Cortical ripples tend to occur at 200 − 500 ms before the peak of spindle power. Moreover, cortical ripples tend to appear at the transition from the DOWN-to-UP states of cortical slow oscillations. Like hippocampal ripples, cortical ripple occurrence is highest during non-REM sleep, intermediate during quiet wakefulness, and low during REM sleep (Khodagholy et al., 2017). The firing rates of most cortical principal neurons and interneurons are positively modulated by cortical ripples (Khodagholy et al., 2017;Miyawaki and Mizuseki, 2022) (Fig. 2) and phase-locked to cortical ripples. In each cycle of cortical ripple oscillations, the firing of principal neurons leads to interneuron firing, similar to the lead-lag hippocampal pyramidal-interneuron patterns during hippocampal ripples (Khodagholy et al., 2017). Furthermore, the strength of the ripple coupling between the hippocampus and posterior parietal cortex during sleep increases after spatial learning (Khodagholy et al., 2017), suggesting that coordinated ripple oscillations play a significant role in the hippocampus-neocortex dialogue and memory consolidation.
In the human brain, cortical ripples in various neocortical areas tend to co-occur with hippocampal ripples (Dickey et al., 2022). Cortical ripples in different neocortical regions are phase-locked, but ripples in the neocortex and hippocampus are not, suggesting that ripple-band cortico-cortical synchrony is mediated by cortico-cortical connections (Dickey et al., 2022). During non-REM sleep, cortical ripples typically occur during the cortical DOWN-to-UP state transition (Dickey et al., 2022). The increase in the co-occurrence of cortico-cortical and hippocampo-cortical ripples precedes successful delayed memory recall of paired-associated words (Dickey et al., 2022), supporting the notion that temporal coordination of ripples in different brain regions is instrumental for memory consolidation and retrieval.

Amygdalar HFOs
Amygdalar HFOs (approximately 200 Hz) (Ponomarenko et al., 2003) (Fig. 1), also known as ripples (Cox et al., 2020;Perumal et al., 2021), occur primarily during non-REM sleep but rarely occur during waking or REM sleep (but see (Haufler and Pare, 2014)). Amygdalar HFOs show a lower number of cycles (3-6 versus 11 cycles) and a smaller amplitude than hippocampal ripples but have an equivalent oscillation frequency (Ponomarenko et al., 2003) (Fig. 3a). Amygdalar HFOs and ventral hippocampal ripples are not phase-coherent (Ponomarenko et al., 2003), but occur temporally in the vicinity (Miyawaki and Mizuseki, 2022) (Fig. 3b, c). BLA neuron firing is phase-locked to the HFO (Ponomarenko et al., 2003) and enhanced firing activity within HFO (Fig. 2). Similar to the hippocampal SWRs, in many cases, the amygdalar HFO is accompanied by a 15-100-ms sharp potential in the 1-20 Hz band (Ponomarenko et al., 2003), suggesting that the HFO is generated by a mechanism similar to the hippocampal SWRs, where the shape wave-associated dendritic depolarization is coupled with oscillatory somatic inhibition in the CA1 (Ylinen et al., 1995). Consistent with this notion, recent ex vivo experiments showed that sharp-wave-coupled HFOs are generated in the local circuit of the BLA; chandelier interneuron discharges, which evoke time-locked feedback excitatory input, can initiate amygdalar HFOs (Perumal et al., 2021).
HFOs coupled with sharp waves are also observed in the human amygdala (Cox et al., 2020). Amygdalar HFOs tend to co-occur with hippocampal SWRs and cortical spindles (Cox et al., 2020), suggesting that amygdalar HFOs are involved in the hippocampus-amygdala-neocortex information transfer and memory consolidation.

Hippocampal, cortical, and amygdalar fast network oscillations coordinate cross-regional ensemble coactivation during memory consolidation and retrieval
As discussed, fast network oscillations during non-REM sleep are observed across various brain regions. The fast network oscillations in the dorsal hippocampus and parietal cortex have been demonstrated to be coupled during sleep, and the coupling is enhanced after spatial learning (Khodagholy et al., 2017). Furthermore, recordings from patients with epilepsy have shown that hippocampal SWRs and neocortical ripples are significantly coupled during successful memory retrieval and that this coupling is accompanied by the reinstatement of memory-specific cortical activity (Norman et al., 2019;Vaz et al., 2019Vaz et al., , 2020. These observations suggest that coupling between hippocampal and neocortical ripples supports memory consolidation and retrieval. Furthermore, a recent study involving approximately 17 h of continuous recording from hundreds of neurons in the amygdala, prefrontal cortex, and ventral hippocampus of freely moving rats revealed that fast network oscillations mediated inter-regional communication changes through cued fear conditioning and the subsequent sleep epochs (Miyawaki and Mizuseki, 2022). In that study, memory-related neuronal ensembles in the amygdala, prefrontal cortex, and ventral hippocampus were determined separately during fear conditioning. The activations of the ensembles were assessed during sleep epochs preceding and following the conditioning sessions. In post-conditioning non-REM sleep epochs, the memory-related ensembles reactivated simultaneously   (Figs. 4, 5). Notably, such coactivations of neuronal ensembles were not detected in pre-conditioning sleep. Furthermore, synchronous reactivation across the three brain regions (the amygdala, prefrontal cortex, and ventral hippocampus) also occurred with fast network oscillations in the involved areas. The triple activation of the ensembles was also enhanced in post-conditioning sleep compared to pre-conditioning sleep. These results indicate that inter-regional interactions via fast network oscillations are involved in fear memory (Miyawaki and Mizuseki, 2022).
Although both amygdalar-prefrontal cortical coactivations and ventral hippocampal-prefrontal cortical coactivations are enhanced through fear conditioning, their time evolutions are diverse across region pairs. The memory-related neuronal ensembles in the amygdala and prefrontal cortex are coactivated during conditioning sessions, especially at the shock onset, and the coactivations persist in the subsequent non-REM sleep periods. In contrast, coactivation between ensembles in the ventral hippocampus and prefrontal cortex is subtle during conditioning but prominent during the subsequent non-REM sleep periods (Miyawaki and Mizuseki, 2022). These observations indicate that reorganization of inter-regional networks occurs during non-REM sleep.
Cortical slow oscillations (Steriade et al., 1993a;, which reflect synchronous alternations between the UP-and DOWN-states across the neuronal population, have been proposed to coordinate inter-regional interactions by aligning fast network oscillations (Klinzing et al., 2019). Consistently, inter-regional ensemble coactivations are also modulated by cortical slow oscillations. Interestingly, amygdalar-prefrontal coactivations occur predominantly at UP-to-DOWN transitions, whereas hippocampal-prefrontal coactivations prefer DOWN-to-UP transitions. Thus, the periods for stabilizing existing networks and developing novel networks may be separated by silent periods of slow oscillations (Miyawaki and Mizuseki, 2022).
The reorganization of the inter-regional network during sleep epochs may shape activity patterns in the subsequent awake periods. Interregional coactivations observed during post-conditioning sleep reemerge during memory retrieval, which is also hosted by fast network oscillations (Fig. 5). However, temporal evolutions of coactivation strengths differ across region pairs. The coactivation of ensembles in the amygdala and prefrontal cortex diminished in retrieval sessions compared with conditioning sessions. On the other hand, the ensemble coactivations between the ventral hippocampus and prefrontal cortex were more robust in retrieval than in conditioning sessions, suggesting that these newly developed ensemble coactivations during sleep are involved in fear memory retrieval; thus, sleep epochs play active roles in memory function (Miyawaki and Mizuseki, 2022).
As discussed above, inter-regional coactivations of local ensembles change through fear-conditioning and subsequent sleep. Changes in the inter-regional network, rather than those within local ensembles, may be more responsible for memory formation. In line with this possibility, the amygdalar and prefrontal cortical ensembles are configured prior to conditioning. In contrast, hippocampal ensembles that participate in inter-regional coactivation develop in an experience-dependent manner. Based on these observations, we hypothesize that individual information regarding an experience is captured rapidly by pre-configured local ensembles. In contrast, experience-dependent inter-regional networks bundle the distributed information together to support memory (Miyawaki and Mizuseki, 2022) (Fig. 6). Our hypothesis also implies that the same or similar activity patterns in local circuits do not necessarily induce the same responses in the downstream area and guide behavior, highlighting the importance of monitoring multiple brain regions simultaneously to understand brain functions.

Summary
• Fast network oscillations, such as hippocampal ripples, cortical ripples, and amygdalar HFOs, are frequently associated with sharp waves reflecting synchronous excitatory inputs and are generated by local circuit mechanisms involving inhibitory interneurons. • Fast network oscillations are modulated by slower oscillations, such as thalamocortical spindles and neocortical slow oscillations, during non-REM sleep. These across-regional cross-frequency couplings may facilitate memory consolidation. • Fast network oscillations in different brain regions are temporally coordinated in various combinations. • Cross-regional couplings of fast network oscillations are enhanced by learning and host inter-regional coactivations of memory-related local ensembles, which presumably support memory consolidation and retrieval during sleep and wakefulness, respectively.

Declaration of interest
None.

Data Availability
Data will be made available on request. Fig. 6. Pre-existing local networks are bound inter-regionally in an experience-dependent manner. Schematic summary of local and interregional dynamics of neuronal ensembles. Ensembles in the amygdala and prefrontal cortex are configured prior to fear conditioning. Ensembles in the ventral hippocampus appear during fear conditioning. In addition, the coactivation of ensembles in the amygdala-prefrontal cortex occurs at shock presentation. In addition to the amygdala-prefrontal cortex coactivation, the ventral hippocampus-prefrontal cortex coactivations develop during the sleep epochs following conditioning sessions. These coactivations during sleep epochs are hosted by fast network oscillations, such as hippocampal sharp-wave ripples (SWRs), amygdalar high-frequency oscillations (HFOs), and prefrontal cortical ripples (cRipples). In the subsequent awake periods during cueretention sessions, the inter-regional coactivation re-emerges with fast network oscillations. Reproduced from the Physiological Society of Japan Science Topics -153 Figure (http://int.physiology.jp/en/st153/) with permission.