Instructive signals drive learning through behavioral timescale synaptic plasticity in the hippocampus and beyond Authors: Lucius K Wilmerding, Fox Gourianova, Christine Grienberger Published: 2026 ABSTRACT AbstractLearning requires the brain to link experience to lasting changes in neural activity, yet most experiences unfold over timescales far longer than those of classical synaptic plasticity mechanisms. Behavioral timescale synaptic plasticity (BTSP) has emerged as a candidate mechanism that bridges this gap by enabling rapid, one-shot modification of synaptic strength at the seconds timescale. BTSP is driven by dendritic Ca2+ plateau potentials (“plateaus”) and potentiates inputs active within a seconds-long window surrounding the plateau, providing a mechanism for the rapid formation and updating of hippocampal place cell representations. Here, we review recent experimental and theoretical advances in BTSP, with a particular focus on the origin and function of the instructive signal from medial entorhinal cortex layer 3 (MEC3) that drives plateau generation in CA1 pyramidal neurons. We discuss evidence that this instructive signal may interact with elements of the hippocampal microcircuit to shape learning-dependent representations. Beyond the hippocampus, emerging findings indicate that BTSP-like mechanisms may operate in other cortical regions, suggesting that integrating signals across dendritic compartments may represent a more general motif for rapid learning. Together, these observations support a view of BTSP as a multi-factor plasticity rule that links synaptic, circuit, and systems-level processes, providing a biologically grounded framework for understanding how the brain rapidly incorporates new information. Introduction Our environment is constantly changing, requiring the brain to continually learn from experience to guide adaptive behavior (Clewett et al., 2019; Holtmaat and Svoboda, 2009). Most experiences unfold over timescales far longer than the firing dynamics of a single neuron. Moreover, sensory inputs are often noisy or unreliable, yet must still support rapid adaptation, sometimes after only a single exposure. An effective learning system must, therefore, enhance salient information for storage while filtering out noise. Together, these constraints pose a fundamental problem for the brain: how to assign credit across time - linking outcomes to relevant, past neural activity - in order to build representations that capture the meaningful structure of experience and support future predictions (Knill and Pouget, 2004; Trommershauser et al., 2011). Decades of research have established the hippocampal-entorhinal circuit as a central system for learning and memory. Early evidence came from the case of patient H.M., whose profound amnesia following medial temporal lobe resection demonstrated the necessity of this region for memory formation (Scoville and Milner, 1957). Subsequent work in animal models, using spatial memory tasks such as the Morris water maze (Morris, 1981; Morris et al., 1982), further established the hippocampus as critical for spatial learning and navigation. Together, these findings position the hippocampal-entorhinal circuit as a key system for investigating the cellular mechanisms underlying learning. In this context, a new form of plasticity, behavioral timescale synaptic plasticity (BTSP), which has recently been identified in mouse hippocampal neurons, offers a potential mechanism linking single experiences and the synaptic modifications required to encode them (Bittner et al., 2017; Dorian et al., 2026; Fan et al., 2023; Grienberger and Magee, 2022; Priestley et al., 2022; Zhao et al., 2020). A central feature of BTSP is the division of labor between distinct dendritic compartments of pyramidal neurons: the apical tuft integrates an instructive target signal driving plasticity, whereas the proximal dendritic compartment contains the synapses that primarily undergo plasticity. In this way, BTSP emerges from the ability of pyramidal neurons to function as multi-compartment - or “two-point” - integrators, providing a cellular basis for the formation of flexible spatial representations that emphasize goal-relevant stimuli or locations and, more generally, experience-dependent representations. Two recent reviews have provided an excellent synthesis of the empirical findings on BTSP (Madar et al., 2025b; Magee, 2026). Here, we take a complementary perspective by focusing on a less explored aspect of BTSP: the origin of the instructive target signal conveyed to apical tuft dendrites. Target signals are theoretically powerful drivers of desired activity states, acting as an external yardstick against which local error computations tune the network's output during learning (Lillicrap et al., 2020). Target signals remain constitutively active during learning, distinguishing them from external feedback error signals, which diminish as the learning network approaches the desired output. Viewed in this context, the target signal proposed in the BTSP working model may offer a generalizable framework for understanding how persistent instructive signals coordinate synaptic plasticity across hierarchically organized brain circuits. We further link our insights from the hippocampus, where BTSP was first identified, to emerging findings in visual cortical regions, which suggest similar principles may underlie learning in neocortical pyramidal neurons more broadly. Finally, we connect these biological learning rules to computational theories in which efficient learning requires the rapid integration of information across multiple sources to extract meaningful structure from large, unstructured datasets. While the focus of this review is on synaptic plasticity as a substrate of learning, we note that other forms of plasticity, including changes in intrinsic excitability and homeostatic forms of plasticity (Debanne et al., 2019; Wen and Turrigiano, 2024), also critically contribute to shaping neural representations during learning. A brief overview of synaptic learning rules Previous pivotal work has established that changes in the strength of synaptic connections between neurons (i.e., synaptic plasticity) are essential for experience-dependent modifications of neural computations during learning (as reviewed in Caya-Bissonnette and Béïque, 2024). One of the most influential concepts in synaptic plasticity research stems from Hebb's postulate that repeated and correlated activity of pre- and postsynaptic neurons strengthens their connection (Fig. 1a; Hebb, 1949). This principle was classically demonstrated by the high-frequency stimulation protocols developed by Timothy Bliss and Terje Lømo to induce long-term potentiation (LTP) (Bliss and Lømo, 1973). While LTP established activity-dependent synaptic strengthening, spike-timing-dependent plasticity (STDP) introduced precise temporal constraints, requiring pre- and postsynaptic activation to occur within tens of milliseconds to induce synaptic change (Fig. 1b; Bi and Poo, 1998; Markram et al., 1997; L. I. Zhang et al., 1998). Hebbian plasticity mechanisms are often associated with unsupervised learning, in which relationships between inputs and outputs emerge without an explicit goal. Such rules are well-suited for extracting correlations in data, useful for processing diverse inputs, and have broadly influenced machine learning approaches that operate on large, unlabeled datasets (Brunel, 1996). Comparison of synaptic plasticity rules. a) Hebbian long-term potentiation (LTP) or depression (LTD) of synaptic weights relies on the repeated coincidence or divergence of action potentials in the pre- (A) and post- (B) synaptic neurons. b) Hebbian STDP leads to LTP or LTD based on the relative timing of the pre- and post-synaptic action potentials after repeated pairings at the millisecond timescale. c) Three-factor plasticity involves repeated pairing of pre- and post-synaptic spikes that lead to a seconds-long eligibility trace. The presence of a third factor gates the transformation of the synapse-specific eligibility trace into synaptic change - either LTP or LTD depending on the circuit and the identity of the third factor. d) BTSP relies on the coincidence of a seconds-long eligibility trace following synaptic input and an apical tuft dendritic input, which leads to a dendritic Ca2+ plateau potential (“plateau”). Eligible synapses are potentiated, whereas some may be depotentiated depending on their initial strength and timing relative to the plateau. However, real-world events and their outcomes can unfold over seconds to minutes, or even longer, posing a challenge for learning rules operating on millisecond timescales. To address this temporal mismatch, theoretical and experimental work converged on three-factor learning rules (Fig. 1c; reviewed in Gerstner et al., 2018), in which coincident pre- and postsynaptic activity produces a transient, synapse-specific eligibility trace that decays over seconds. This trace does not immediately alter synaptic strength: instead, it marks synapses as eligible for modification. A subsequent third factor, often a neuromodulator or another signal that conveys information about reward, salience, or surprise, for example, converts this latent trace into a lasting change in synaptic weight. Experimental support for three-factor rules driving LTP or long-term depression (LTD) depending on the specific neuromodulator has emerged across multiple brain regions, including cortex and hippocampus (Brzosko et al., 2015, 2017; He et al., 2015). A particularly clear example comes from the striatum, where pairing pre- and postsynaptic activity induces an eligibility trace that can be transformed into synaptic potentiation by a delayed dopamine signal within a narrow temporal window of approximately 1 s (Yagishita et al., 2014). Outside this window, no lasting plasticity occurs, indicating that synaptic change depends on the interaction between a transient eligibility trace and a temporally precise third factor. While three-factor rules do extend Hebbian plasticity to behaviorally relevant timescales, they still generally require repeated pairings to produce eligibility traces, have not been extensively validated in vivo, and do not explicitly incorporate feedback signals about behavioral outcomes in ways that support rapid learning from single experiences. Thus, an important question remains: how can the brain support such learning? In the following sections, we consider BTSP, a form of plasticity that can be induced in a single trial, operates on behavioral timescales, and enables neuronal populations to refine their activity toward a target pattern (Fig. 1d). The hippocampus as a case study for BTSP BTSP was first identified in the hippocampus (Bittner et al., 2015, 2017), a region long studied for its role in learning, memory, and spatial navigation (Barnes, 1979; Marr, 1971; McNaughton and Morris, 1987; Morris et al., 1982; Scoville and Milner, 1957). Subsequent work across multiple laboratories has supported and extended the initial findings on BTSP (Fan et al., 2023; Madar et al., 2025a; Priestley et al., 2022; Zhao et al., 2021). Pyramidal neurons in the hippocampal CA1 subregion exhibit spatially selective firing patterns (Fig. 2a) (C. D. Harvey et al., 2009; Moser et al., 2015; O'Keefe and Dostrovsky, 1971). Populations of these so-called place cells tile the environment and are thought to support the formation of cognitive maps as long-lasting representations of spatial experiences (R. Muller and Kubie, 1987; O'Keefe and Nadel, 1978; Pastalkova et al., 2008; Tolman, 1948). In non-uniform environments, place fields tend to cluster around salient features, such as rewards or predictive cues (Fig. 2b) (Dupret et al., 2010; Grienberger and Magee, 2022; Hollup et al., 2001; Kaufman et al., 2020). Thus, a mechanistic understanding of place field formation may be critical to unlocking deeper insight into the encoding of spatial, and, as we discuss later, potentially non-spatial, episodic memories. Characteristics of BTSP in hippocampal CA1. a) Silent CA1 pyramidal neurons do not exhibit place field firing as an animal traverses a track (left), as the membrane voltage (Vm) remains below the action potential threshold. A plateau (middle) consisting of a sustained depolarization (blue) and action potentials (black) causes synaptic weight modifications. As a result, a place field (right) predictive of the plateau location is formed. b) Place cells tile the entire environment (top). Adding a salient cue or reward (green, bottom) leads to an overrepresentation of these locations through BTSP. c) CA1 pyramidal neurons receive excitatory synaptic input from medial entorhinal cortex layer 3 (MEC3) onto their apical tuft and from CA3 onto their proximal dendrites, which have been proposed to provide the instructive target signal and drive the eligibility trace, respectively. Stratum lacunosum moleculare (SLM), parvalbumin-positive (PV), vasoactive intestinal polypeptide-expressing (VIP), bistratified cells (BiC), and oriens lacunosum moleculare (OLM) interneurons are known to synapse with each other and/or with CA1 neurons. Their precise function in regulating BTSP is an active area of research. d) Distal tuft-targeted MEC input drives plateaus associated with a global intracellular Ca2+ increase (red). NMDA receptor and CaV2.3 voltage-gated Ca2+ channel antagonists block plateau initiation. A seconds-long local eligibility trace (yellow) outlasts the post-synaptic response generated by upstream CA3 input to the proximal dendritic compartment. The overlap between these two signals is thought to underlie the BTSP time course and associated synapse potentiation. Cortical pyramidal neurons segregate sensory feedforward inputs from long-range feedback inputs to different dendritic compartments, which, when coincident, lead to non-linear, Ca2+-dependent signaling of coincident inputs – a theoretical hallmark of associative learning (M. Larkum, 2013; M. Larkum et al., 1999). Similarly, CA1 pyramidal neurons receive their main excitatory synaptic inputs from hippocampal area CA3 and from entorhinal cortex layer 3 (EC3), with these inputs being spatially segregated, targeting proximal (i.e., basal and apical oblique) and apical tuft dendrites, respectively (Fig. 2c) (Kajiwara et al., 2008; Steward and Scoville, 1976; Takács et al., 2012). More specifically, contralateral CA3 primarily targets CA1 basal dendrites in stratum oriens, while ipsilateral CA3 axons tend to synapse on the apical oblique dendrites in stratum radiatum (Tamamaki et al., 1988). Across the cortex and hippocampus, basal, apical oblique, and apical tuft dendrites can support compartmentalized computations through local spine and/or dendritic Ca2+ signals involving NMDA receptors (Kovalchuk et al., 2000; M. E. Larkum et al., 2009; Nakamura et al., 2002; Schiller et al., 2000). Consistent with this organization, CA1 pyramidal dendrites can undergo local plasticity (Frick et al., 2004; Kim et al., 2007), enabling compartment-specific computation and storage (Losonczy et al., 2008), although the degree of separation between somatic and branch activity may be small in vivo and dependent on the behavioral state of the animal (Lee et al., 2026; X. Wu et al., 2026). Within this spatially organized input architecture, convergent EC3 input, supported by proximal CA3 input, initiates the dendritic Ca2+ potential (“plateau”), the key event driving BTSP in CA1 (Fig. 2a) (Bittner et al., 2015; Grienberger and Magee, 2022; Takahashi and Magee, 2009). Plateau events depend on NMDA receptor activation and are characterized by sustained depolarizations mediated by voltage-gated Ca2+ channels, lasting tens to hundreds of milliseconds and reflected somatically as so-called complex spike bursts (Fig. 2d) (Epsztein et al., 2011; Grienberger et al., 2014; Takahashi and Magee, 2009). One or only a few such plateaus are sufficient to induce BTSP, leading to synaptic weight changes primarily in those CA3-to-CA1 inputs that are active within a seconds-long window surrounding plateau initiation (Bittner et al., 2017) and providing a potential mechanism by which a single experience can affect synaptic weight. In the context of spatial navigation, when CA3 inputs are themselves spatially tuned (Gonzalez et al., 2025), this process converts a previously silent CA1 pyramidal neuron into a spatially selective place cell. CA3 inputs to CA1 apical oblique dendrites constitute the most likely population of synapses modified by BTSP, while CA3 inputs to CA1 basal dendrites may also undergo plasticity (Fan et al., 2023; but see Gonzalez et al., 2025). Notably, BTSP can be induced experimentally in both silent neurons and pre-existing place cells by triggering plateaus via current injection, juxtacellular electrical stimulation, or optogenetic activation in vivo (Bittner et al., 2015, 2017; Diamantaki et al., 2018; Fan et al., 2023; Gonzalez et al., 2025; Grienberger et al., 2017; Milstein et al., 2021; O'Hare et al., 2022; Rolotti et al., 2022). BTSP in CA1 exhibits an asymmetric, seconds-long plasticity kernel that preferentially strengthens inputs preceding the plateau (Bittner et al., 2017). As a result, place fields generated by BTSP are shifted backward relative to the location of plateau induction (“predictive place fields”) and vary in width as a function of running speed - broader in regions where animals move faster and narrower where movement is slower (Bittner et al., 2017; Milstein et al., 2021; Priestley et al., 2022). The molecular mechanism underlying the time course of this plasticity kernel remains under active investigation. Computational modeling suggests that it could arise from the interaction between a local eligibility trace and a plateau-associated global signal (Fig. 2d). Here, somewhat reminiscent of the three-factor learning rules but without requiring postsynaptic action potential activity, presynaptic CA3 activity transiently primes specific synapses without immediately inducing plasticity, while the plateau acts as a secondary global signal that converts this latent state into a lasting synaptic modification. The eligibility trace is thought to decay over seconds, thereby playing a key role in shaping the asymmetric time course of BTSP. The molecular basis of the eligibility trace, which serves as a biochemical memory of recent activation, remains unresolved. Given that NMDA receptor-dependent dendritic and/or spine Ca2+ transients can last several hundred milliseconds, longer than the initial EPSP, this Ca2+ signal may play a role in the synaptic eligibility trace, although the time course still falls short of the plasticity window observed in BTSP. Although CaMKII has been implicated (Jain et al., 2024; Xiao et al., 2023), its role in shaping this process remains to be established. Similarly, the mechanistic underpinnings underlying the plateau-associated global signal remain unclear. Evidence points to Ca2+-induced Ca2+ release (CICR) as a contributor to this signal, as blocking Ca2+ sequestration prolongs the BTSP in the forward direction (O'Hare et al., 2022). Additionally, an in vitro study in prefrontal layer 5 pyramidal neurons has suggested that CICR may contribute to both the local and global signals involved in BTSP (Caya-Bissonnette et al., 2023). While the formation of new place fields is critical for navigation, spatial representations must remain plastic to flexibly adapt to new information. BTSP is also involved in reshaping existing place fields in response to changes in salient features (Cone and Shouval, 2021; Diamantaki et al., 2018; Milstein et al., 2021; Qian et al., 2025), and has been implicated in updating spatial representations during a single learning episode (Madar et al., 2025a; Priestley et al., 2022). Inducing BTSP outside an existing field can drive field relocation, even in the absence of external environmental changes, indicating that BTSP is sufficient to reorganize spatial representations (Diamantaki et al., 2018; Milstein et al., 2021). As these neurons typically lose their old field, this implies a bidirectional plasticity kernel that both potentiates and depotentiates synapses within a defined temporal window around the plateau, effectively replacing old fields with new ones (Fig. 3a) (Milstein et al., 2021; Vaidya et al., 2025). In this way, BTSP incorporates a form of intrinsic normalization that stabilizes network representations while preserving flexibility. Specifically, weak or strong synapses active within a few seconds of the plateau potentiate or remain strong, respectively. Strong synapses active at more distant time points undergo depotentiation, and inactive synapses remain unchanged (Fig. 3a; Milstein et al., 2021). Without data on the molecular underpinnings of BTSP and whether they are similar to Hebbian LTP/LTD, we have elected not to label BTSP-mediated synaptic changes as either LTP or LTD. In any case, computational models employing BTSP-like rules for rapid potentiation of correlated inputs and depotentiation (or repulsion) of similar memory patterns exhibit efficient, sparse coding that enables pattern separation and efficient storage (Y. Li et al., 2024; Milstein et al., 2021; Y. Wu and Maass, 2025). Over time, this dual mechanism enables refinement of spatial maps and contributes to the maintenance and reactivation of place fields across days (Vaidya et al., 2025). Entorhinal and hippocampal connectivity supporting an instructive signal for experience-dependent plasticity. a) CA1 pyramidal cells integrate spatially tuned CA3 inputs and MEC3 inputs carrying the instructive signal. Plateau initiation triggers asymmetric updating of CA3 input weights (red: plasticity kernel), resulting in a backward shift of place field activity on subsequent traversals. In existing place fields, some synapses also undergo depotentiation depending on their temporal distance from the plateau and their pre-induction weight. b) CA3 pyramidal cells exhibit symmetric BTSP-mediated weight changes of recurrent inputs from other CA3 cells. c) The most relevant circuit elements are shown. Sensory inputs arrive at MEC2/3 principal cells. MEC2 axons synapse in the dentate gyrus and CA3 tuft dendrites. CA1 and subiculum outputs converge onto deep MEC5, providing multi-synaptic feedback to superficial layers 2 and 3 of the MEC. This feedback may provide a readout of the status of CA1 representations and dictate the ongoing need for the MEC3 target signal. Local inhibitory microcircuits in CA1 are depicted, highlighting their importance in regulating BTSP. Abbreviations match Fig. 2. Taken together, CA1 provides a powerful model system for understanding how this form of non-Hebbian synaptic plasticity can rapidly establish and reorganize neuronal tuning. The seconds-long BTSP plasticity kernel enables the binding of temporally dispersed CA3-to-CA1 inputs at behaviorally relevant timescales, and MEC3 input plays a central role in initiating plateaus. In the following sections, we examine these two excitatory input pathways in greater detail to better understand their distinct contributions to BTSP. An instructive target signal from the medial entorhinal cortex layer 3 The EC is the main gateway between cortical areas and the hippocampus (Fig. 3c) (Insausti et al., 1997; Kerr et al., 2007; Moser et al., 2015). It can be divided into its medial (MEC) and lateral (LEC) components, the former relating more strongly to positional information and the latter to objects, attention, and motivation (Eichenbaum et al., 2012; Witter et al., 2017). Notably, EC3 input to CA1 is organized along a proximodistal gradient: proximal CA1 receives stronger input from MEC3, whereas LEC3 more strongly influences distal CA1 (Bowler and Losonczy, 2023). This anatomical organization suggests that distinct entorhinal streams may differentially shape neural activity along the CA1 axis. To date, most studies of BTSP have focused on the MEC; therefore, we focus our discussion on this region. Within the BTSP framework, the MEC3 supplies an instructive target signal that drives plateaus in CA1 neurons and, as a result, plays a significant role in shaping the structure of CA1 population activity. When particularly salient environmental cues or rewards are introduced, CA1 increases the density of place fields near these prominent features as learning progresses (Fig. 2b; Grienberger and Magee, 2022). The formation of these overrepresentations requires MEC3 activity, as ipsilateral optogenetic silencing of MEC3 abolishes the enrichment while leaving ongoing place field activity intact. Importantly, MEC3 activity is present from the onset of learning, whereas the emergence of the CA1 representation is delayed and appears to follow the MEC3 activity pattern. In addition, MEC3 output remains relatively stable while the rate of BTSP-mediated CA1 place field formation decreases over the course of learning. Together, these findings indicate that 1) the MEC3 signal is more consistent with a target signal than a long-range global error signal, which would decrease with learning, and 2) that local circuit mechanisms, such as feedback inhibition (Campbell et al., 2026; Rolotti et al., 2022), may be key regulators of CA1 BTSP induction, specifically affecting the rate of BTSP while the target signal remains constant (see the section on “Circuit mechanisms influencing BTSP in CA1”). In this view, MEC3 provides an instructive target signal that interacts with the CA1 microcircuit to shape learning-dependent CA1 representations. Yet, how this target signal is computed remains unclear. In the following paragraphs, we examine the intrinsic and extrinsic anatomy of the MEC, its single-cell properties, and its population activity patterns related to behavior, which together may underpin its role in providing the target signal to CA1. Highly processed sensory information is thought to converge onto neurons in the superficial MEC layers 2 and 3, which integrate this input and convey it to all subdivisions of the hippocampus. While MEC2 neurons project primarily to the dentate gyrus (DG) and hippocampal areas CA3 and CA2 by the perforant path, MEC3 neurons project selectively to CA1 and the subiculum by the temporoammonic path (Fig. 3c) (Kajiwara et al., 2008; Steward and Scoville, 1976; Suh et al., 2011; Takács et al., 2012). These distinct projection patterns give rise to a bifurcated information flow from MEC2 and MEC3 that converges in CA1, where processed CA3 outputs meet MEC3 inputs on the same pyramidal neurons. MEC3 neurons further differ from MEC2 cells in their morphology, electrophysiology, and intra-layer connectivity (Witter et al., 2017). Specifically, MEC2 contains two types of excitatory cells, stellate and pyramidal cells, whereas MEC3 is composed predominantly of pyramidal neurons (Tukker et al., 2022). Stellate cells make up approximately 50-75% of the layer 2 excitatory neuron population and project to DG and CA3, while MEC2 pyramids cluster anatomically into so-called “islands” and project to a variety of CA1 interneurons and many ipsilateral and contralateral MEC targets (Fuchs et al., 2016; Kitamura et al., 2014; Rowland et al., 2018; C. Varga et al., 2010). MEC3 pyramidal cell dendrites appear to avoid patches of MEC2 pyramidal cells and preferentially receive input from other MEC3 pyramidal cells as well as MEC2 stellate cells (further reviewed in Tukker et al., 2022). In sum, primarily MEC3 pyramidal cells constitute the source of the target signal. Closing the entorhinal-hippocampal loop, CA1 projects to the subiculum, and both regions send afferents to MEC5b, a functional subpopulation of layer 5. MEC5b neurons, in turn, project to the superficial layers of the MEC, whereas MEC5a cells target other cortical and subcortical regions (Osanai et al., 2023; Sürmeli et al., 2015). MEC5a and MEC5b pyramidal neurons also exhibit distinct physiological properties, and while MEC5b cells innervate MEC5a cells, this connection is weaker in MEC than in LEC (Ohara et al., 2018, 2021). CA1 and subiculum also project to MEC6, which innervates all hippocampal subregions and may play a previously underappreciated role in spatial information processing and memory (Ben-Simon et al., 2022). Recent work in mice has also demonstrated a relatively smaller axonal projection from CA1 and subiculum directly to MEC3 (Butola et al., 2025), consistent with earlier studies in rats and cats (Cenquizca and Swanson, 2007; Ino et al., 2001). Feedback from area CA1 to MEC3, either indirectly through MEC5 or directly, would be well positioned to evaluate the impact of an instructive signal (Lillicrap et al., 2020), such as the MEC3 target signal that drives BTSP. As discussed above, Ca2+ imaging data suggest the MEC3 input to CA1 remains relatively stable near salient cues throughout learning, although some refinement occurs (Grienberger and Magee, 2022), consistent with its acting as a target rather than an error signal. However, how this apparent stability is reconciled with changing CA1-to-MEC feedback remains unclear. One possibility is that this feedback contributes to the observed refinement of the MEC3 target signal. This question is particularly intriguing given evidence that this feedback is functionally important: silencing it impairs the encoding of new memories in hippocampus-dependent contextual fear conditioning (Roy et al., 2017) and novel object recognition (Butola et al., 2025). CA1-to-MEC3 feedback input, which targets the proximal dendrites of MEC3 neurons, can also heterosynaptically potentiate synaptic responses evoked in the distal tuft of the same neurons in vitro (Butola et al., 2025). Causal manipulations of CA1 feedback to distinct MEC layers, combined with simultaneous recordings in CA1 or cellular-resolution measurements across MEC layers during spatial behavior, are currently lacking and would provide valuable mechanistic insight. Specific functional subsets of MEC neurons may be central to generating the target signal. A wide range of spatially tuned cell types associated with allocentric, or map-based, navigation have been described in MEC, including grid cells, speed cells, border cells, head direction cells, and aperiodic spatial cells (Hafting et al., 2008; Hinman et al., 2019; Kropff et al., 2015; Sargolini et al., 2006; Solstad et al., 2008; S.-J. Zhang et al., 2013). Given that grid cells exhibit highly regular and stable firing patterns, their output may be less well suited to account for the relatively intermittent instructive signal described by Grienberger and Magee (2022). Modeling work further suggests that infrequent plateau generation enhances memory sparsity, allowing the network to store more information, consistent with an intermittent instructive signal (Y. Wu and Maass, 2025). Therefore, selective enhancement of CA1 representations at behaviorally salient locations may instead be supported by MEC “cue” cells with firing fields aligned to prominent landmarks, as well as by aperiodic spatial cells and border cells (Kinkhabwala et al., 2020; Solstad et al., 2008; S.-J. Zhang et al., 2013). However, there remains a relative paucity of models and recordings of non-grid cell populations in the MEC, particularly in deeper layers, leaving key questions about their roles in learning and circuit feedback unresolved. The contribution of specific cell types to the instructive signal is likely difficult to disentangle. Manipulations of parvalbumin- and somatostatin-expressing interneurons have been shown to differentially affect grid cells and aperiodic spatial cells, respectively (Miao et al., 2017), offering experimentalists some traction. Experimental designs incorporating more complex behavioral paradigms will be necessary to provide further insight into the nature of the MEC3 target signal. To date, BTSP studies have largely relied on one-dimensional environments with a single reward or predictive cue (Bittner et al., 2015, 2017; Grienberger and Magee, 2022; Y. Li et al., 2024; Milstein et al., 2021). However, one study used a virtual reality paradigm to unexpectedly shift or duplicate visual cues that had previously served as sites of BTSP and place field formation in CA1 (Zhao et al., 2020). Place cells in this context were strongly driven by sensory cues, yet remained highly context-dependent, suggesting that CA1 integrates contextual information from upstream regions to generate a specific spatial code. In experiments in which the reward location was shifted relative to stable environmental cues, BTSP signatures were observed in both CA1 place cells that remained cue-driven and those that remapped to follow the reward (Qian et al., 2025), implying a potential role for both cue and reward selectivity within the target signal. A two-choice virtual reality task also showed that CA1 inherits context-dependent place field activity from upstream regions, primarily CA3 (Zhao et al., 2021). Notably, MEC neurons can also exhibit context-dependent spatial tuning (Lipton et al., 2007), although it remains unclear whether this context dependence originates within MEC or is inherited from upstream regions. Specifically, prefrontal regions, particularly the ventromedial prefrontal cortex, may direct the MEC-hippocampal circuit to perform context disambiguation, especially at remote time points (de Sousa et al., 2026; Preston and Eichenbaum, 2013). Furthermore, interactions between the LEC and CA3 are important for context-association learning (Robert et al., 2025), indicating that both EC subdivisions can contribute to context learning. Therefore, it is conceivable that the MEC target signal itself is subject to top-down contextual modulation, which could bias the active hippocampus-projecting neuronal population and allow contextually separable representations in downstream CA1. Future studies combining multi-region, population-level recordings with tasks involving multiple cues or context switches will be necessary to determine how context affects the MEC3 target signal. An unbiased map from CA3 The hippocampal CA3 region plays a central role in models of episodic memory and pattern completion (Kesner and Rolls, 2015; Marr, 1971; McNaughton and Morris, 1987). CA3 lesions produce errors in the behavioral expression of memory for memorized locations when cues are attenuated, an effect linked to NMDA receptor expression in the recurrent collaterals (Gold and Kesner, 2005; Nakazawa et al., 2002). In the BTSP context, CA3-to-CA1 synapses represent the primary locus of plasticity, making CA3 activity a critical determinant of how experience is encoded in CA1. Here, we focus on anatomical, physiological, and behavioral features of CA3 that are most relevant for understanding how BTSP operates. CA3 receives input from MEC2 stellate cells and from sparse, powerful (“detonator”) synapses originating in the dentate gyrus (Amaral et al., 2007; Kitamura et al., 2014; Squire et al., 1989), so that CA3 activity can reflect both cortical sensory information and pattern-separated representations. A defining feature of CA3 is its extensive recurrent connectivity, with collateral axons forming synapses with other CA3 pyramidal neurons at an estimated probability of ∼4% (Amaral and Witter, 1989; Ishizuka et al., 1990; Treves and Rolls, 1994). This architecture supports auto-associative dynamics that stabilize activity patterns and enable pattern completion (Marr, 1971; McNaughton and Morris, 1987; Treves and Rolls, 1994). From the perspective of BTSP, such stability provides a structured and relatively consistent input to CA1, which the MEC instructive signal can selectively modify. Accordingly, CA3 pyramidal neurons exhibit spatially tuned firing in both one- and two-dimensional environments, often with narrower place fields and higher spatial information than CA1 neurons (Mizuseki et al., 2012). Moreover, CA3 representations are comparatively resistant to gradual changes in environmental cues, maintaining stable firing patterns where other regions remap (Leutgeb et al., 2007; Neunuebel and Knierim, 2014), especially during re-exposure to familiar environments (Dong et al., 2021), although see (Hainmueller and Bartos, 2018). This stability suggests that CA3 provides a relatively fixed spatial scaffold on which BTSP can act in CA1 to generate flexible, experience-dependent representations. Recent in vivo work directly links BTSP to the formation of this spatial scaffold in CA3 (Y. Li et al., 2024). Using whole-cell recordings in behaving mice, CA3 pyramidal neurons were shown to generate plateaus and undergo BTSP for place field induction. However, in contrast to the asymmetric plasticity kernel observed in CA1, CA3 exhibited a symmetric plasticity profile, with comparable depolarization levels before and after the onset of the plateau (Fig. 3b). This symmetry was observed for both natural and experimentally induced place fields. In subsequent experiments, optogenetic inhibition of CA3 recurrent inputs was applied immediately before the experimentally evoked plateau in the recorded CA3 pyramidal cells. This manipulation led to a strong reduction in plasticity at the inhibited location immediately before the induction location, but not after. Inhibiting the EC had the opposite effect, possibly because EC activity is required to update the active CA3 ensemble and, in turn, the inputs to the recorded CA3 cell. DG inhibition had minimal effect, indicating that CA3 and EC are the primary contributors to this process, at least under these experimental conditions. These experiments were performed in mice already familiar with the environment, and the observation of spontaneous place field induction therefore suggests that refinement and novel field formation continue in CA3 even in experienced animals. The authors also developed a CA3 model incorporating a BTSP learning rule, which rapidly adjusted synaptic weights and generated uniformly distributed place fields within a few laps. This finding suggests that, during learning in a novel environment, BTSP may quickly establish a spatial scaffold in CA3. In vivo experiments will be needed to validate this prediction. Together, these findings suggest that CA3 exhibits BTSP but operates under different plasticity rules than CA1, potentially reflecting distinct functional roles within the circuit. Computational modeling of CA3 supports this division of labor. Network models incorporating BTSP demonstrate that reversible potentiation and depotentiation can prevent saturation in auto-associative networks and improve the separability of overlapping input patterns (Y. Li et al., 2024; Y. Wu and Maass, 2025). Models also emphasize that CA3 provides a spatially stable, attractor-like representation, whereas CA1, via asymmetric BTSP, transforms this input into predictive representations that emphasize behaviorally relevant features. Taken together, these findings support a model in which CA3 provides a stable, spatially structured input that defines the substrate for plasticity, while BTSP in CA1 selectively modifies this input in response to instructive signals. In this way, CA3 and CA1 form a complementary system in which stable representations are flexibly reshaped to encode experience. Circuit mechanisms influencing BTSP in CA1 Although CA3 and MEC3 are key elements of current BTSP models, CA1 microcircuit components also play an important role in regulating plateau occurrence and BTSP induction. CA1 pyramidal neurons exhibit substantial heterogeneity, which may support diverse microcircuit computations and downstream outputs (Geiller et al., 2017; R. E. Harvey et al., 2023; Soltesz and Losonczy, 2018). In parallel, interneurons shape hippocampal output (Kepecs and Fishell, 2014; Klausberger and Somogyi, 2008), with multiple subtypes contributing to remapping in novel environments and the enrichment of representations at goal locations during learning (Geiller et al., 2020). Different interneuron classes are well-positioned to influence BTSP by targeting complementary compartments of pyramidal neurons (Fig. 2c). Accordingly, inhibiting hippocampal interneurons broadly increases place field induction (Rolotti et al., 2022). Soma-targeting parvalbumin-positive (PV) basket cells are strongly regulated by behavioral state and modulate the dynamic range and participation of pyramidal neurons in circuit computations (Geiller et al., 2020; Hu et al., 2014). They are well positioned to influence BTSP induction, although precisely how remains to be determined. Bistratified cells (BiCs) target the apical oblique and basal dendrites of CA1 pyramidal neurons, where excitatory inputs from CA3 form synapses, and their activity is modulated by novelty, running velocity, and proximity to goals (Geiller et al., 2020; Halasy et al., 1996; Klausberger et al., 2004). These properties position BiCs to regulate incoming CA3 input and, potentially, the resulting eligibility traces at these synapses. Consistent with this possibility, exposure to a novel environment transiently suppresses the activity of somatostatin-positive dendrite-targeting interneurons, including BiCs (Sheffield et al., 2017). Apical tuft-targeting interneurons, such as oriens–lacunosum moleculare (OLM) and stratum lacunosum moleculare (SLM) interneurons, may more directly influence plateau generation (see below). Vasoactive intestinal polypeptide-expressing (VIP) interneurons, in turn, act in a disinhibitory manner by suppressing other interneuron populations, including OLM and basket cells (Turi et al., 2019), and may be preferentially engaged during novelty to transiently open windows for plasticity, allowing for place field formation (Neubrandt et al., 2025). Finally, long-range excitatory and inhibitory inputs further regulate interneuron activity (Basu et al., 2013; Hernández-Frausto et al., 2023), potentially fine-tuning the timing of plateau initiation during behavior. The role of apical tuft region-targeting interneurons in regulating plateau generation, especially OLMs, is under active investigation. OLMs exhibit bidirectional connectivity with CA1 pyramidal neurons, placing them in a position to provide feedback inhibition of ongoing hippocampal activity on a network level (Fig. 2c; Fig. 3c; Chamberland et al., 2024). Recent work by Campbell et al. (2026) found that activation of a genetically defined subset of OLM neurons (Leão et al., 2012) suppresses BTSP-mediated formation of new place fields early in learning, whereas silencing these neurons later in learning promotes additional place field formation. These results suggest that OLM interneurons regulate BTSP during learning. Specifically, increased activity of OLM inhibitory interneurons later in learning suppresses BTSP, preventing additional place fields from forming despite a constant target signal from the MEC. Mechanistically, this is consistent with OLM-mediated inhibition of apical tuft dendrites (Chamberland et al., 2024; Milstein et al., 2015; Udakis et al., 2025) and with a role for these interneurons in counteracting the MEC3-derived instructive signal. Learning-dependent increases in OLM feedback may progressively limit the influence of this signal (Campbell et al., 2026; Sheffield et al., 2017), thereby constraining further plasticity once representations have been established, while preserving the flexibility needed to incorporate environmental changes into the representation. This interpretation is also consistent with a disinhibitory effect of VIP interneurons on pyramidal neurons via their inhibition of OLM interneurons, especially during novelty when BTSP-mediated place field formation is high (Neubrandt et al., 2025; Turi et al., 2019). SLM interneurons also target the apical tuft, and themselves integrate input from MEC3 as well as thalamic and cortical sources, providing an additional potential pathway through which extrahippocampal signals can influence CA1 representations (Capogna, 2011; Freund and Buzsáki, 1998; Maccaferri, 2011). Together, these observations support a view of BTSP as a multi-factor plasticity rule in which dendritic inhibition and disinhibition dynamically regulate when and where instructive signals can drive synaptic change. Another important microcircuit element is neuromodulation. The hippocampus receives dopaminergic input from the ventral tegmental area (VTA), particularly targeting the stratum oriens in CA1 (Adeniyi et al., 2020; Adeyelu and Ogundele, 2023; Takeuchi et al., 2016). VTA activity and dopamine release bias CA1 place cell populations toward salient locations, such as those associated with reward (Krishnan et al., 2022; Mamad et al., 2017). In addition, dopaminergic and noradrenergic input to CA1 arises from the locus coeruleus (LC) (Smith and Greene, 2012; Takeuchi et al., 2016). While VTA axonal activity ramps up as animals approach expected reward locations late in learning (Krishnan et al., 2022), LC axonal activity remains relatively constant in familiar environments but increases markedly during context switches, consistent with a novelty signal (Heer and Sheffield, 2024). Manipulating LC activity can modulate place cell overrepresentation, enhancing responses to familiar rewards or reducing responses to novel rewards (Kaufman et al., 2020). Together, these findings suggest that catecholamines – dopamine and norepinephrine – play an important role in stabilizing and enhancing spatial representations of salient stimuli in the hippocampus. Acetylcholine is also hypothesized to play an important role in hippocampal memory encoding and retrieval (Hasselmo, 2006; Hasselmo et al., 1995) and modulates the strength of tetanic LTP induction in vitro and in vivo (Adams et al., 2004; Huerta and Lisman, 1995; Ovsepian et al., 2004). Furthermore, acetylcholine facilitates and prolongs somatic and dendritic Ca2+ transients evoked by excitatory input (Giessel and Sabatini, 2010; Kis et al., 2024), which could help initiate or sustain plateaus driving BTSP. A direct link between neuromodulatory signaling and BTSP remains to be established, but dopamine, norepinephrine, acetylcholine, and serotonin act as gating factors for LTP or LTD in vitro in three-factor plasticity rules across the hippocampus, cortex, and striatum (Brzosko et al., 2015, 2017; He et al., 2015; Yagishita et al., 2014), offering an important direction for future BTSP research. In sum, both local inhibition and neuromodulatory input are likely to provide additional control over BTSP by modulating the probability of plateau initiation. Computational models emphasize the importance of plateau frequency (Y. Wu and Maass, 2025) for memory sparsity and storage capacity. During novelty or proximity to salient features, neuromodulators such as dopamine and norepinephrine may increase the probability of a plateau. To balance the increased probability, increases in OLM interneuron activity and reduced disinhibition from VIP interneurons later in the learning episode may reduce the occurrence of plateaus, thereby helping maintain sparsity and stability in hippocampal representations. Finally, extensive work has shown that high-firing-rate neurons are more likely to be incorporated into engram, or memory trace, ensembles in regions such as the lateral amygdala and DG (Frankland and Josselyn, 2015; Sekeres et al., 2012; Yiu et al., 2014; Zhou et al., 2009). Given that CA1 pyramidal neurons with higher firing rates exhibit an increased probability of plateau generation (Bittner et al., 2015), this raises the possibility that BTSP may contribute to engram formation beyond spatial contexts, including associative learning paradigms such as fear conditioning. Relationship of BTSP and theta rhythm Oscillatory electric field activity, or brain rhythms, is thought to coordinate the activity of many neurons, supporting complex behaviors such as learning. Although a full understanding of the interaction between BTSP and the theta rhythm has yet to be established, relevant findings point toward a relationship between these two phenomena. In the hippocampus, the theta rhythm (7-12 Hz in rodents) strongly modulates the activity of pyramidal neurons and interneurons, likely organizing computations into discrete packets (Colgin, 2013; Lisman and Idiart, 1995). GABAergic neurons from the subcortical medial septum project throughout the hippocampus and the EC, and act as a global pacemaker to entrain local unit activity to the theta frequency (Buzsáki et al., 1983; Freund and Antal, 1988; Mizumori et al., 1990; Tóth et al., 1997; V. Varga et al., 2008). Theta power is elevated during active exploration in rodents, marking a shift to an "online" state associated with the encoding of new information (Terrazas et al., 2005; Vanderwolf, 1969). Consistent with this view, the two major inputs to CA1, MEC3 and CA3, are temporally segregated across distinct theta cycle phases (Brankačk et al., 1993; Hasselmo et al., 2002; Hölscher et al., 1997; Mizuseki et al., 2009) (Fig. 4a), suggesting a structured temporal organization for information routing. Against this backdrop, the precise timing of spikes relative to the extracellular theta phase has led to the concept of phase coding, which may offer computational advantages beyond those of rate-based coding alone (Jensen and Lisman, 2000; Lisman, 1999). Plateau occurrence in relation to theta rhythmic activity in CA1. a) Timing of hippocampal CA1 inputs and CA1 neuronal activity relative to theta rhythm. b) Schematic of theta phase precession in a CA1 pyramidal neuron as a mouse moves through the cell's place field. Spikes at place field entry occur late relative to theta, while spikes at field exit occur early. PV - Parvalbumin, OLM - Oriens Lacunosum Moleculare, BiC – Bistratified cell. Neurons in all hippocampal subregions and the EC exhibit a phenomenon known as theta phase precession (Hafting et al., 2008; O'Keefe and Recce, 1993; Skaggs et al., 1996). Briefly, spikes early in a place field occur at later phases of the theta cycle, while spikes later in the field shift to earlier phases (Fig. 4b) (O'Keefe and Recce, 1993; Skaggs et al., 1996). Place cells with spatially close fields often exhibit tighter co-firing relationships, called theta sequences (Dragoi and Buzsáki, 2006; Foster and Wilson, 2007; Gupta et al., 2012; Wikenheiser and Redish, 2015). Inputs from the MEC are required for this temporal coding strategy in CA1 neurons, as manipulations of MEC activity disrupt normal phase procession (Guardamagna et al., 2025; Schlesiger et al., 2015), although precession can still be stimulated in CA1 in the absence of precessing inputs (Sloin et al., 2024). In addition, the MEC and CA1 exhibit interregional coupling at theta frequency (Zhou et al., 2022; Zutshi et al., 2021). The importance of phase coding for hippocampal place field activity raises the possibility of an interaction between theta and BTSP. CA1 plateaus tend to cluster at or just after the peak of the theta oscillation measured in the CA1 pyramidal cell layer, a phase during which both MEC3 and CA3 inputs are active (Hasselmo et al., 2002; Mizuseki et al., 2009) (Fig. 4a). Accordingly, neurons firing at earlier theta phases exhibit a higher probability of plateau generation (Bittner et al., 2015). The observation that plateaus do not cluster in the late phase of theta, despite a similar balance of MEC3 and CA3 inputs as in the early phase of the theta cycle (Fig. 4a), points to additional, potentially local circuit mechanisms that regulate the timing of plateau initiation and the selective recruitment of specific CA1 neuron subsets for BTSP-mediated place field formation during novel experience. Further investigation into the relationship between theta phase, plateau induction, and the temporal sequences of newly formed place cells may yield considerable insight into the temporal aspects of BTSP and memory formation. Lastly, during non-exploratory behavior, the hippocampus enters a distinct state characterized by low theta amplitude and the appearance of transient, high-frequency oscillations called sharp wave ripples (SPW-Rs), a signature of memory consolidation (Buzsáki, 2015; Ego-Stengel and Wilson, 2007; Girardeau et al., 2009; Roux et al., 2017). SPW-Rs coordinate activity between ensembles of task-activated neurons in the hippocampus and neocortex, including the visual cortex during sleep (Ji and Wilson, 2007), and the medial prefrontal cortex during awake learning (Jadhav et al., 2016; Shin et al., 2019). Notably, following BTSP-mediated place field induction, CA1 pyramidal neurons exhibit greater membrane potential depolarization and action potential firing during SPW-Rs than before induction (Bittner et al., 2015). Neurons that have undergone BTSP may therefore be preferentially recruited into ensemble activity associated with memory consolidation during offline, low-theta states. Generalizing BTSP BTSP provides many potential computational advantages for the brain and for brain-inspired networks. First, the asymmetric shape of the plasticity kernel in the CA1 results in many place fields shifting backward relative to the plateau location. Therefore, the new field becomes a prospective representation of the original location that drove the plateau (Bittner et al., 2017; Milstein et al., 2021), in line with theories and findings that the hippocampus creates a predictive map rather than a purely spatial one (Stachenfeld et al., 2017; Yaghoubi et al., 2026). Along the same lines, BTSP in CA1 is sufficient to generate successor-like representations, in which neural activity reflects the expected future occupancy of states rather than purely instantaneous position. This predictive structure emerges naturally from the asymmetric plasticity kernel, linking synaptic dynamics to forward-looking representations (Galloni et al., 2024). In contrast, the symmetric kernel found in CA3 creates distinct “attractors” or noise-resistant activity states of a dynamical system (Fig. 3a and b) (Kesner and Rolls, 2015; Khona and Fiete, 2022), cued by external MEC2 input but largely supported by the recurrent architecture of this region (Y. Li et al., 2024). Second, modeling work supports BTSP's ability to create rapid, sparse memories that can be recalled with partial input cues (P. Y. Li and Roxin, 2023; Y. Li et al., 2024; Y. Wu and Maass, 2025). The speed of this encoding is critical for successful navigation, enabling the brain to readily map cue representations onto neural space without repeatedly traversing the environment. In this sense, BTSP enables efficient content-addressable memory by leveraging sparse coding to conserve energy and encode a vast number of “units” of data, whether in spatial code or in computing (P. Y. Li and Roxin, 2023; Y. Li et al., 2024; Y. Wu and Maass, 2025). Maps of different environments or contexts can be modeled as attractors supported by the recurrent connectivity of CA3, and adaptively used by CA1 to support remapping without overburdening the system's storage capacity (P. Y. Li and Roxin, 2023; Milstein et al., 2021). Content-addressable memory (CAM) is a desirable search feature in computing that maps data to a location based on its content. Computational models have achieved CAM using a BTSP rule and simple binary weights (P. Y. Li and Roxin, 2023; Y. Wu and Maass, 2025), rather than continuous weights, which were necessary in Hopfield networks to achieve high-capacity CAM (Hopfield, 1982). Given that high-capacity, sparse CAM representations can be formed with this simple learning rule and binary weights, translation to hardware applications may be feasible (Y. Wu and Maass, 2025). A recent model incorporated multi-compartment neurons and an MEC-like target signal to train an artificial neural network, demonstrating the feasibility of biologically inspired credit assignment for image classification (Galloni et al., 2026). Plateaus may likewise be used as a general tool to create region-specific effects. A recent study examined the properties of plateaus in the subiculum (Bhatia et al., 2026), a major hippocampal output region innervated primarily by CA1 and EC3. The authors used whole-cell patch-clamp recording in head-fixed mice navigating a linear treadmill. They found that subiculum neurons exhibited plateaus at a much higher rate than CA1 neurons, with a substantial fraction clustering near reward locations, potentially to transmit information about these salient events beyond the hippocampus (Bhatia et al., 2026; Simonnet and Brecht, 2019). Given that similar two-compartment dendritic and input architectures are found across many brain regions (M. Larkum, 2013), including the subicular circuit, it is likely that plateaus serve distinct functions depending on region and cell type. Although much of our previous discussion has concerned spatial representations in hippocampal circuits, plateaus and BTSP may not be limited to place field formation. To date, experimental place field induction has required that the animal be engaged in running behavior with strong theta-rhythmic activity (Bittner et al., 2017; Diamantaki et al., 2018; Y. Li et al., 2024; Zhao et al., 2020). A recent Ca2+ imaging study investigated the role of BTSP in the rapid formation of odor-tuned representations in CA1 pyramidal cells that rely on the entorhinal cortex, although this technique precluded analysis of local field potential activity to determine whether high-theta power was required for successful odor field induction (Dorian et al., 2026). Thus, plateaus may also be initiated during other periods of high theta activity, such as REM sleep (Buzsáki, 2002; Jouvet, 1969). Given evidence for pre-existing sequential activity that precedes behavioral experience (Dragoi and Tonegawa, 2011, 2013), it is intriguing to consider the role of plateaus in biasing ensemble formation during sleep. An experiment investigating CA3 or CA1 during REM sleep would likely require voltage imaging to fully dissect theta rhythmicity, plateaus, and increased co-activity between cell populations, but would provide valuable data toward understanding whether BTSP can function in states other than active running behavior. Finally, BTSP-like mechanisms may extend beyond the hippocampus to other brain regions. Two recent preprints report activity-dependent modifications following plateau firing in layer V neurons of a higher visual area (Xiao et al., 2025) and in primary visual cortex (Yaeger et al., 2025). These findings indicate that the integration of signals across dendritic compartments may represent a general prerequisite for BTSP-like plasticity throughout the brain, particularly in cortical pyramidal neurons with apical tuft dendrites that express the ion channels required for plateau generation (M. E. Larkum et al., 2009). As in CA1 and CA3, long-duration plateaus initiated one- or few-shot plasticity in these neocortical cells. The change in membrane voltage also exhibited a bidirectional plasticity kernel, further underscoring the concurrent potentiation and depotentiation as key features of BTSP (Xiao et al., 2025). However, some differences from the hippocampus were noted, including a generally lower magnitude of potentiation (Xiao et al., 2025) and failure of plateaus to alter basic response properties, such as orientation selectivity (Yaeger et al., 2025), suggesting that some inputs are less responsive to BTSP mechanisms. Taken together, these findings raise the possibility that BTSP represents a broadly conserved form of supervised plasticity, which enables rapid learning across cortical regions while preserving a more stable, “hardwired” circuit structure. Modulatory influences are also likely to shape BTSP beyond the hippocampus. Cellular and circuit-level factors, including differences among principal cell subtypes, the activity of diverse interneuron populations, and neuromodulatory inputs, are expected to influence the probability of plateau generation and, consequently, the rate of plasticity induction. Finally, we hypothesize that BTSP is best suited for circuits that support rapid and flexible learning. The hippocampus is a clear example, as patterns of neural activity must often be encoded after a single experience to form long-lasting episodic memories. The prefrontal cortex supports inference, allowing existing knowledge to be applied to novel situations and adaptive behavior to emerge without the need for potentially costly direct experience (Abdou et al., 2024; Brunamonti et al., 2016; DeVito et al., 2010; Preston and Eichenbaum, 2013; Ramawat et al., 2022). In this context, BTSP could enable prefrontal neurons to bind information acquired separately and support transitive inference across novel combinations of stimuli. As discussed above, plateaus also operate in deep layers of visual cortical regions to enhance responses to behaviorally relevant stimuli (Xiao et al., 2025; Yaeger et al., 2025). In contrast, BTSP may be less appropriate in more hardwired circuits, such as subcortical sensory pathways or circuits close to motor output, where large and rapid changes in synaptic weight could be maladaptive. Similarly, in the absence of a target signal specifying a desired activity pattern and driving plateau generation, slower correlation-based Hebbian plasticity rules may be better suited to extracting useful output representations from unstructured inputs. Counterpoint: pre-existing circuit architecture The above-mentioned studies of BTSP, including those that use “artificial” (e.g., current injection) place field induction, support an experience-dependent model of hippocampal function. In this view, relatively uniform CA3 inputs converge onto a CA1 pyramidal neuron, while an entorhinal instructive signal drives plateaus that potentiate active synapses, thereby forming novel place fields. Importantly, neurons can, in principle, acquire place fields at any location in the environment, provided they receive appropriately tuned CA3 input there (Fan et al., 2023; Gonzalez et al., 2025). A distinct body of literature argues that external input is instead mapped onto pre-existing hippocampal weights and hard-wired firing relationships between pairs of cells or ensembles (Buzsáki and Tingley, 2018). CA1 pyramidal cells fire in sequence during sleep in advance of exposure to novel environments, where they subsequently exhibit matching sequential place field activity, in a phenomenon termed ‘preplay’ (Dragoi and Tonegawa, 2011, 2013; Grosmark and Buzsaki, 2016). Additionally, in the absence of sensory input, ensembles of CA1 pyramidal cells self-organize into stereotyped sequences that match voluntary run trajectories (Villette et al., 2015). A recent study investigated place field induction in small groups of neurons using focal optogenetic activation of CA1 pyramidal neurons (McKenzie et al., 2021). While stimulation at a fixed location across serial trials on a linear track caused CA1 place fields to remap, these fields did not cluster near the stimulation zone, as predicted by theories of arbitrary external mapping onto the hippocampus. Instead, weak activity observed before stimulation predicted the eventual remapping of the field. These findings suggested that pre-existing circuit dynamics constrain both the emergence of place fields and the co-firing relationships among CA1 neurons during in vivo learning. Another study used precise optogenetic stimulation to induce place field formation in single pyramidal neurons, leading to disinhibition by local interneurons and even recruiting other local pyramidal cells to become spatially tuned themselves (Geiller et al., 2021). These findings, suggesting inherent CA1 connectivity, are supported by a study showing that CA1 neurons with similar birthdates exhibit similar spatial firing patterns across distinct environments, suggesting in utero pre-configuration of hippocampal circuits (Huszár et al., 2022). It is reasonable to hypothesize that BTSP manipulations of single place cells lead to similar local circuit effects, whereas stimulation of larger groups of CA1 pyramidal cells is subject to feedback inhibition, which controls the population-level outcomes and maintains sparser representations (Rolotti et al., 2022). Synthesizing these two bodies of literature suggests that individual place fields can be induced at essentially any location in an environment via BTSP, either artificially or by a strong MEC3 drive. Once established, these fields may disinhibit nearby pyramidal neurons through pre-existing local interneuron connectivity, allowing them to acquire similar tuning (Geiller et al., 2021). At the population level, pre-existing circuit architecture and fluctuations in neuronal excitability (Frankland and Josselyn, 2015) may bias certain groups of pyramidal neurons to undergo BTSP and develop place fields. The size of the recruited ensemble and the rate of place field induction are likely further regulated by diverse interneuron populations and neuromodulatory signals. How these pre-existing circuit constraints interact with the more flexible mechanisms of BTSP to shape the formation of new CA1 place fields during actual learning remains an important open question. Counterpoint: alternative plasticity rules An alternative class of models proposes that STDP, in combination with theta rhythmicity, can support spatial representations in the hippocampus, particularly in CA3 (Bush et al., 2010; George et al., 2023). For example, it was proposed that the cognitive map formed in CA1 is maintained by Hebbian plasticity between the recurrently connected cells of CA3 (R. U. Muller and Stead, 1996). Modeling the confluence of inputs from grid cells with overlapping fields leads to unique hippocampal place codes, consistent with Hebbian plasticity operating on millisecond timescales (Savelli and Knierim, 2010). In simulations, an STDP rule was also found to produce biologically realistic CA3 activity when combined with a timing restriction rule corresponding to the length of a theta cycle, demonstrating that the temporal control provided by theta rhythmicity may be sufficient to account for key features of hippocampal firing patterns (Bush et al., 2010). STDP models have also been used to account for changes in pre-existing CA1 place fields, with the asymmetry of the STDP kernel leading to skewed fields after repeated traversals (Mehta et al., 2000). More recent modeling work demonstrated that STDP can create successor representations in CA1 from spatial activity in CA3, using the temporal compression of theta sweeps to sustain ordered representations (George et al., 2023). Multiple plasticity rules likely contribute to the formation and refinement of hippocampal memory representations. BTSP provides an in vivo-validated synaptic plasticity rule that can account for several prominent features of hippocampal activity during learning. These include one-shot learning, integration of inputs across seconds, large backward shifts of place fields from the place field induction trial to the next trial, the adaptive overrepresentation of salient environmental features, and sparse, non-overlapping coding for efficient memory storage - many of which are difficult to reconcile with STDP or even three-factor plasticity rules alone. Hebbian millisecond-timescale plasticity rules may operate in the hippocampus to maintain and refine spatial representations and sequential activity within theta cycles, and to interact with SPW-R-related memory consolidation. These could also serve to stabilize a subset of “important” synapses against natural mechanisms of synaptic turnover. Neuromodulators released after animals experience the outcome of an action may further induce synaptic weight changes, depending on the neuromodulator and circuit, although this prediction has yet to be validated in vivo. Conclusions and outlook Insights from the hippocampal-entorhinal circuit offer explanations for how the brain transforms noisy inputs into robust, organized representations. Each subregion serves as both an independent computational unit and a component of a larger system that integrates diverse signals to form useful memories. Within this framework, BTSP provides a compelling mechanism for combining semi-independent sources of information to generate representations from single learning experiences. By operating on longer timescales than classical plasticity rules, BTSP offers a more direct link between real-world experience and circuit-level computation, with important implications for understanding behavior and for the development of biologically inspired learning algorithms. We anticipate that continued investigation of memory through the lens of BTSP will advance both neuroscientific and computational approaches to learning. Many questions regarding BTSP remain open. For example, the role of neuromodulatory systems, such as dopaminergic and noradrenergic inputs, in regulating plateau probability and shaping instructive signals remains to be fully understood. In addition, elucidating the molecular mechanisms underlying BTSP will be essential for bridging circuit-level function with the cellular and biochemical processes that give rise to learning. Furthermore, while BTSP provides a compelling single-cell mechanism for connecting single experiences to synaptic change, it likely operates alongside other forms of synaptic plasticity and circuit mechanisms that contribute to learning across different timescales and brain regions. Unraveling how these mechanisms interact remains an important question. Understanding how memory representations form at the circuit level is also essential for addressing cognitive decline associated with aging and neurodegenerative disease. Both normal aging and pathological neurodegeneration are accompanied by structural atrophy and alterations in neural activity, as extensively reviewed elsewhere (Hess, 2005; Jahn, 2013). Rodent models of Alzheimer's disease, the most common cause of dementia in elderly populations (Götz et al., 2018; Hendrie, 1998), provide valuable insight into the mechanisms of neurodegeneration and potential therapeutic strategies by enabling access to neural signaling at cellular and molecular levels (Braidy et al., 2012). CA1 exhibits impaired place field remapping in a knock-in model of Alzheimer's disease, alongside impairments in spatial tuning in the MEC (Jun et al., 2020), indicating a specific deficit in the cognitive map's adaptability. Given the role of BTSP in shaping memory representations, disruptions to the entorhinal instructive signal during aging and disease represent an important avenue for future investigation. A greater understanding of the plasticity mechanisms underlying the hippocampal-entorhinal circuit may also inform the development of data-driven computational systems. BTSP-like learning rules show promise for unsupervised learning in feedforward competitive networks (Cone et al., 2025), and emerging work suggests that they may enhance the speed and robustness of hierarchical models capable of binding and unbinding higher-order representations (Yu et al., 2025) and performing image classification (Galloni et al., 2026). More broadly, the hippocampal-entorhinal circuit provides a powerful example of how structured representations can be extracted from relatively unstructured sensory and reward signals, an ability that is central to many applications in artificial intelligence (Richards et al., 2019), and critical to the understanding of the brain. Credit author statement Lucius K Wilmerding: Conceptualization, Writing-Original Draft, Writing-Review&Editing, Visualization. Fox Gourianova: Conceptualization, Writing-Original Draft, Writing-Review&Editing, Visualization. Christine Grienberger: Conceptualization, Writing-Original Draft, Writing-Review&Editing, Visualization, Supervision. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements This work was supported by the National Institutes of Health 1DP2 MH136393 (CG), 4RF1 MH135576 (CG), T90 DA059112 (FG), F32 NS143954-01 (LKW), Air Force Office of Scientific Research FA 9550-22-1-0344 (CG), the Smith Foundation (CG), the Pew Foundation (CG), the Alfred P. Sloan Foundation (CG), and the McKnight Neuroscience Scholar Award (CG). Footnotes This article is part of a special issue entitled: Two-Point Neurons published in Current Research in Neurobiology. Data availability No data was used for the research described in the article. References Associated Data Data Availability Statement No data was used for the research described in the article.