A fundamental challenge for memory systems is balancing two opposing demands: sensitivity, which enables the encoding of subtle differences between similar experiences, and consistency, which preserves stable representations against interference. How the brain resolves this trade-off has remained unclear. Here, we identify inhibition in the dentate gyrus (DG) as a regulator of this balance, shifting hippocampal computation between sensitivity- and consistency-dominated regimes. Using an integrative approach that combines cell-type–specific pharmacogenetics in mice, behavioral assays, and computational modeling, we demonstrate that reducing inhibition during the encoding of overlapping experiences enhances later discrimination during recall. Computational modeling further predicted that, besides discrimination enhancement, disinhibition also increases susceptibility to interference, an effect that becomes particularly critical under high memory load, and compromises memory consistency across experiences. Conversely, increased inhibition stabilized memory representations, but reduced sensitivity, rendering similar experiences indistinguishable. These specific predictions were confirmed experimentally, showing that the DG inhibition tunes a sensitivity-consistency continuum. Together, our results uncover a systems-level principle by which hippocampal circuits dynamically prioritize flexibility or stability in memory formation depending on inhibitory tone. This framework advances our understanding of how the brain balances competing computational demands and has implications for neuropsychiatric and neurodegenerative disorders in which excitation–inhibition balance is disrupted.

DG PV⁺ cells regulate GCs output bidirectionally

We regulated PV-cell activity in the DG with Designer Receptors Exclusively Activated by Designer Drugs (DREADDs) [17]. These receptors allow gain- or loss-of-function manipulations of PV-cell firing, depending on the subtype of receptor expressed, without imposing an external pattern of activation nor completely blocking spiking. We injected stereotaxically adeno-associated virus 5 (AAV5)-human synapsin1 (hSyn)-DIO-hM4D(Gi)-mCherry into the hilus of the DG of PV-cre transgenic mice (129P2-Pvalbtm1(cre)Arbr/J) to express the inhibitory hM4D(Gi) receptor in dentate PV⁺ interneurons (Fig 1a; referred from now as PV-Gi animals). Conversely, for the activation of PV-cells we injected the AAV5-hSyn-DIO-hM3D(Gq)-mCherry virus to express the excitatory hM3D(Gq) receptor (Fig 1a; referred to as PV-Gq animals). The proportion of PV⁺ interneurons infected in the DG was 92.8 ± 3.8% and 91.2 ± 3.5% for hM4D(Gi) or hM3D(Gq) viruses, respectively (S1a–S1c Fig).

Electrophysiological recordings were performed in vivo before and 90 min after Clozapine-N-Oxide (CNO) administration (i.p. 1 mg/Kg) and again 60 min following the induction of long-term potentiation (LTP). In PV-Gi animals, evoked electrophysiological potentials in the DG following stimulation of the perforant pathway (PP), the main entorhinal cortex input to the hippocampus, showed a significant increase in the amplitude of the population spike (PS) after CNO (i.p. 1 mg/Kg) administration (Fig 1b), reflecting the facilitated firing of GCs. Conversely, decreased PS amplitude in the DG (Fig 1b) was recorded in PV-Gq animals after CNO administration, indicative of the hindered GC firing. Importantly, the excitatory postsynaptic potentials (EPSPs), reflecting the synaptic responses and dendritic integration in GCs, were unchanged in all groups (Fig 1c). The integrity of the dendritic function was further demonstrated in synaptic plasticity experiments in which the induction of LTP by high-frequency stimulation of the PP increased the EPSP in all groups (Fig 1d). Consistent with this, no significant differences in the magnitude of the EPSP after LTP were observed between experimental groups (see S1 Table).

Although viral expression was not strictly restricted to the DG, infected PV⁺ neurons were predominantly located in this region, with lower levels of infection observed in CA3 and minimal infection in CA1 (DG: 93 ± 3%; CA3: 33 ± 5%; CA1: 5 ± 2%). Thus, the manipulation primarily targeted DG PV⁺ interneurons, although a contribution from CA3 PV⁺ interneurons cannot be excluded. To assess cellular specificity, we quantified the proportion of DREADD-expressing cells that co-expressed PV and found that 96.96 ± 1.13% of infected cells were PV⁺, indicating highly selective targeting of the intended interneuron population.

To assess the functional impact of this manipulation on hippocampal circuit activity, we compared responses through the trisynaptic pathway (EC → DG → CA3 → CA1) and the disynaptic pathway bypassing the DG (EC → CA3 → CA1) in anesthetized mice (S2 Fig). PV⁺ interneuron manipulation altered activity propagation through the trisynaptic pathway, whereas transmission through the disynaptic pathway remained largely unchanged. This pathway-specific effect supports the interpretation that the manipulation mainly affected DG-dependent hippocampal processing, rather than producing a generalized alteration of CA3-driven transmission to CA1.

DG PV⁺ cells modulation alters spatial discrimination

We then investigated whether the modulation of GC’s output, while preserving intact synaptic and dendritic function, had any impact on memory formation. To this end, we used the hippocampal-dependent novel object location task (NOL) [18,19]. In this task (Fig 2a), mice first explore a context (habituation) and are then reintroduced to the same context containing two identical objects during an exploration session (familiarization). After 24 h, they are re-exposed to the same context, but with one object displaced (test). If spatial memory is intact, animals show a preference for the object that occupies a novel position (see Materials and methods and Fig 2a).

To assess the role of DG PV⁺ interneurons during distinct stages of memory processing, we administered CNO (i.p. 1 mg/Kg) either 90 min before encoding, 10 min after encoding (consolidation), or 90 min before retrieval (Fig 2a). Behavioral performance, measured as a discrimination index (DI; see Materials and methods), was significantly influenced by PV⁺ cell activity modulation during memory encoding (Fig 2b), but not during consolidation (Fig 2c) or retrieval (Fig 2d). Reducing perisomatic inhibition (PV-Gi) during encoding improved memory performance 24 h later (Fig 2b), whereas increasing PV⁺ interneuron activity (PV-Gq) impaired it (Fig 2b). Importantly, total object exploration time during the encoding session did not differ between groups (Sham: 19.98 ± 0.08 s; PV-Gi: 19.41 ± 0.61 s; PV-Gq: 18.94 ± 0.79 s; Kruskal-Wallis test: H(2) = 1.073, p = 0.584), indicating comparable exposure to the objects during memory acquisition. These memory-related effects were not attributable to differences in locomotor activity or anxiety-like behavior, as assessed by open-field exploration and the Elevated Plus Maze (EPM) (S3b–S3e Fig). As a critical control, the NOL task was repeated in the same group of animals, substituting CNO with its vehicle (saline) during encoding. Under these conditions, all groups showed comparable memory performance (Fig 2e), confirming that the observed effects were specific to DREADDs activation.

The impaired performance observed in PV-Gq animals raised the question of whether increased PV-mediated inhibition produced a general deficit in spatial memory formation or, alternatively, a more specific reduction in discrimination capacity. To address this issue, we performed an additional NOL experiment using a larger object displacement (20 cm). In this condition, all experimental groups, including PV-Gq animals, successfully discriminated the displaced object (Fig 2f), indicating that spatial memory formation remained intact in PV-Gq mice when discrimination demands were reduced. These results suggest that the increase of PV-mediated inhibition during encoding does not impair the ability to form spatial memories per se.

To validate the effectiveness of DREADDs-mediated modulation of PV⁺ interneurons during behavior, we stained c-Fos protein and quantified c-Fos/PV co-labeling in the DG following the familiarization phase of the NOL task (S3a Fig). This analysis revealed a significant reduction in the number of PV⁺ cells expressing c-Fos in PV-Gi animals and a significant increase in PV-Gq animals following memory encoding (S3b Fig), consistent with effective pharmacogenetic modulation. Additionally, the overall efficacy of the task to activate cell assemblies in the DG was confirmed by a higher number of c-Fos⁺ cells after NOL exploration compared to home-cage controls (S2c, S2d Fig).

We next asked whether the improved memory performance observed with reduced PV⁺ cell activity came at the expense of spatial discrimination, a function thought to benefit from the sparse activation of GCs [4,5]. To address this, we used a sequential NOL task (see Materials and methods), in which animals were exposed to progressively more challenging spatial changes: an initial test with a subtle displacement (10 cm), followed by a second test with a larger displacement (15 cm). PV⁺ cell activity reduction enhanced fine spatial discrimination, particularly under the most challenging condition (Fig 2g – 2h). When the displacement was small (10 cm), PV-Gi animals injected with saline failed to detect the change, while the same animals under CNO reliably identified it, indicating increased capacity to discriminate subtle spatial changes. When the displacement was increased to 15 cm, under both conditions, animals successfully detected the change.

Together, these results suggest that perisomatic inhibition via PV⁺ interneurons influences the precision of spatial encoding, with reduced inhibition enhancing discrimination. Importantly, the impaired performance of PV-Gq animals at the 15 cm displacement, together with their successful discrimination at 20 cm, supports the interpretation that increased perisomatic inhibition does not prevent spatial memory formation per se, but rather limits spatial discrimination when the task requires finer resolution. Furthermore, these results raised an important question about the broader consequences of modulating inhibition: if a lower inhibitory tone improves performance, why doesn’t the system operate in that state by default?

Computational insights into inhibition-mediated memory dynamics

We developed a cortico-hippocampal neural network designed to emulate an animal’s decision-making process during exploration of a familiar environment, where objects may be encountered in familiar or novel locations. In this model, the level of novelty in a “sensory” input pattern determines whether the model generates a response to explore the NOL. The model consists of two modules (Fig 3a), designed to capture key aspects of hippocampal and prefrontal cortex function: the hippocampal-related module implements Hebbian learning and discrimination, while the prefrontal-related module implements decision-making.

To simulate PV-mediated inhibitory control within the hippocampal network, we applied a k%-max winner-take-all rule in the hippocampal-related module [9] (see Materials and methods). The k% parameter determines the neurons that are allowed to fire: higher k% values correspond to lower inhibition, facilitating neuronal activation, while lower k% values indicate stronger inhibition, limiting neuronal activation. For each inhibition level, the model was trained with a predefined set of “sensory” input patterns, representing activity from a third neuronal layer. It was then tested with gradually modified versions of those patterns to simulate increasing novelty (see Materials and methods). The network’s ability to maintain stable memory representations depended critically on inhibition (Fig 3b). Under high inhibition (low k%), representations in the hippocampal layer remained stable across a wider range of input pattern changes, suggesting that consistency was preserved at the cost of reduced sensitivity. In contrast, under low inhibition (high k%), even small changes in input produced substantial alterations in output representations, indicating increased sensitivity at the cost of consistency. This trade-off was reflected in the steeper decline of output similarity (measured as correlation) for similar inputs under low inhibition (Fig 3b).

To connect these representational changes to behavior, we projected the hippocampal output to the prefrontal-related module, where the model simulated exploration decisions based on the familiarity of incoming sensory patterns. We computed a DI to quantify the model’s tendency to explore objects versus the arena, depending on input novelty (see Materials and methods). As suggested by the hippocampal patterns, the relationship between input similarity and discrimination depended strongly on inhibition level: DI values were higher under low inhibition, especially for subtly modified inputs, reflecting greater sensitivity to change (Fig 3c). This sensitivity was further quantified by comparing DI across inhibition levels for specific novelty conditions (Fig 3d), revealing a marked enhancement in novelty detection at low inhibition (high k%) when input similarity was high (0.8). At lower input similarity (0.6), discrimination was less sensitive to disinhibition, and higher inhibition levels still permitted successful discrimination (Fig 3d). These simulations closely mirrored our experimental findings. When we selected two inhibition levels to reflect control and PV-Gi conditions (intermediate and low inhibition, respectively), the model reproduced the discrimination effect observed in the sequential NOL task (compare Figs 2h and 3e).

These results suggest that reduced inhibition improves discrimination by amplifying the networks’ responsiveness to subtle input changes. However, this increased sensitivity may come at a cost. In biologically realistic networks subject to noise and ongoing plasticity, excessive sensitivity could compromise network consistency. To examine this possibility, we introduced synaptic interference into the model by allowing stochastic fluctuations in synaptic weights. This manipulation caused greater disruption under low inhibition, as reflected by an increased Euclidean distance between the memory representations evoked by the same input before and after interference (Fig 3f). In contrast, high inhibition preserved more stable representations across conditions. These findings highlight the critical role of inhibition in preserving representational consistency and raise an important question: does reduced inhibition compromise memory stability by increasing susceptibility to synaptic plasticity-induced changes?

To investigate this, we systematically increased the number of learned patterns in the network and calculated the DI for each condition. Under low inhibition, novelty-driven exploration declined sharply with increased memory load, eventually reaching chance levels (DI ≈ 0.5), indicating that familiar and novel inputs could no longer be distinguished (Fig 3g). In contrast, higher levels of inhibition preserved discriminability across memory loads. However, this came at the cost of sensitivity: very high inhibition levels were also associated with reduced DI, likely because excessive inhibitory tone produced memory representations that were robust to input changes. As a result, the network’s ability to form sufficiently differentiated representations for similar patterns was diminished, leading to reduced exploration of novel stimuli. These effects are further illustrated in Fig 3h, which shows DI across inhibition levels for a fixed, high memory load. The resulting bell-shaped curve indicates that both low and high inhibition levels impair discrimination in this regime. Together, these results reveal a crucial trade-off: while low inhibition supports finer discrimination, it compromises memory stability by increasing susceptibility to interference. High inhibition, conversely, protects against interference, but limits the network’s sensitivity to subtle changes. These computational predictions were directly tested in a subsequent behavioral experiment.

Modulation of PV⁺ interneurons alters memory interference

To experimentally validate the model predictions, we designed a new behavioral protocol that we called multi-arena NOL task. This task was specifically designed to assess how manipulations of PV⁺ inhibitory activity affect memory interference in conditions of high memory load. The multi-arena NOL task evaluates an animal’s ability to maintain multiple distinct memories through four unique contexts, each defined by multiple environmental cues (color, shape, and landmarks), different objects, and specific object-context-location associations (Fig 4a). The complete experimental protocol was conducted over three consecutive days. Analysis of the averaged DI (aDI) revealed significant group differences (Fig 4b). Notably, PV-Gi animals – where inhibition was reduced and discrimination enhanced in low memory-load experiments (Fig 3b, 3g) – showed in this task a pronounced impairment compared to controls, whereas PV-Gq animals, with enhanced inhibition, exhibited intermediate performance. These behavioral results mirror the model’s predictions: insufficient inhibition increases memory interference, while high inhibition supports more stable memory traces at the cost of discriminability, thereby impacting animals’ behavior.

We further estimated the number of contexts successfully encoded by each animal. This was done by comparing individual DI ratios for each context against a threshold derived from variability in nonlearning conditions (Fig 4c; see Materials and methods). The number of contexts learned varied across groups, with a significant reduction in PV-Gi animals compared to Sham controls, and a trend toward reduction, but at intermediate levels, in PV-Gq animals (Fig 4d).

Together, these findings support the idea that, as memory load increases, maintaining PV⁺-mediated inhibition within a functional range is critical for preserving the fidelity of episodic memory representations. Departure from this balance—either through excessive disinhibition or hyperinhibition—compromises the system’s ability to store and retrieve multiple distinct memories.

Overall, the convergence of computational and experimental results suggests that PV⁺ interneurons in the DG regulate the position of the system along a sensitivity-consistency continuum of memory representation (Fig 5).

Circuit dynamics

Episodic memories are initially acquired within the hippocampus–entorhinal cortex network through rapid synaptic plasticity and are subsequently consolidated in neocortical regions for long-term storage [5,23-29]. The interaction between the hippocampus and the prefrontal cortex plays a crucial role during the encoding of new memories and is essential for their later recall [30]. Modulation of hippocampal inhibitory circuits, as performed in our study—particularly during the encoding phase—is therefore expected to influence this broader memory network. Indeed, we have previously shown, using brain-wide imaging [31], that activity modulation in the DG [32–34] and CA3 [35] regulates functional connectivity across a wide network of memory-related structures, including the hippocampal formation, the prefrontal cortex, and the nucleus accumbens.

While our results show that reducing PV-mediated inhibition enhances fine spatial discrimination, an important question is how such modulation may be flexibly achieved under physiological conditions in the intact DG. PV interneuron activity is embedded within a broader inhibitory network that includes somatostatin (SOM)-expressing interneurons and disinhibitory VIP-expressing cells, whose coordinated activity is dynamically regulated by behavioral state and local circuit interactions [36, 37]. In addition, PV⁺ interneurons dynamics are influenced by neuromodulatory inputs and activity-dependent plasticity within inhibitory circuits [38–40]. Recent work suggests that novelty processing in the DG is associated with transient suppression of PV⁺ interneuron activity [15], supporting the idea that inhibitory tone in DG can be flexibly adjusted during memory encoding. In this context, recent findings showing that dentate spikes support offline hippocampal reactivation and flexible object-location memory [41] raise the possibility that transient DG network states, potentially involving rapid shifts in inhibitory dynamics, contribute to the encoding and discrimination of highly similar spatial representations. Such transient reductions in PV-mediated inhibition could provide a physiological mechanism for temporarily shifting the DG toward a high-sensitivity state, thereby facilitating the separation of overlapping experiences when fine discrimination is required. Conversely, restoration of inhibitory tone may help preserve representational consistency once memories have been encoded.

Although the present study focused on novelty-based object-location memory, the proposed mechanism is unlikely to be restricted to novelty processing. DG-dependent reference memory tasks requiring fine spatial discrimination also depend on the separation of overlapping spatial representations. We therefore expect PV-mediated inhibition to regulate a similar sensitivity-consistency balance in those tasks, although the upstream signals controlling inhibitory tone may differ from those engaged during novelty detection.

Computational model

Our computational model was designed to capture essential principles of how inhibition shapes memory computations within a cortico-hippocampal system, specifically the trade-off between sensitivity and consistency in the context of novelty-driven decision-making. Rather than aiming for detailed biological realism, we prioritized conceptual clarity and generalizability. Moreover, the model does not explicitly represent task-specific variables such as contextual identity, object identity, arena geometry, or phase structure, which differ across the behavioral paradigms employed experimentally. Instead, memories are encoded as distributed representations without explicit contextual labels or procedural features. This level of abstraction allowed us to isolate the effects of inhibitory tone on representational dynamics and memory-guided behavior without relying on complex parameter tuning or intricate circuit assumptions that could reduce interpretability. While more detailed models, such as biophysically realistic spiking networks, can capture finer aspects of memory and decision-making [42–44], they often involve fitting numerous parameters, which can limit their explanatory power and increase the risk of overfitting [45]. By contrast, our model balances between simplicity and explanatory depth. Conceptual models of this kind have been widely and successfully used to uncover the mechanisms behind memory and decision-making [46–49]. Our model provides a minimal yet insightful computational framework for understanding how inhibitory modulation may influence memory-based computations, offering a plausible mechanistic perspective on systems-level memory function.

Methodological considerations

A methodological consideration is that the viral manipulation was not strictly confined to the DG and also affected PV⁺ interneurons in CA3. Thus, our results should not be interpreted as arising from an exclusively DG-restricted manipulation. Nevertheless, this limitation does not alter the main conclusion of the study. Anatomically, viral expression was strongly predominant in the DG, with lower infection in CA3 and minimal infection in CA1. Functionally, PV⁺ interneuron manipulation altered activity propagation through the DG-dependent trisynaptic pathway, whereas transmission through the disynaptic pathway bypassing the DG remained largely unchanged. These observations support the interpretation that the behavioral effects are primarily linked to altered DG-dependent processing, while acknowledging that a contribution from CA3 PV⁺ interneurons cannot be fully excluded. Importantly, the central concept proposed here does not require an absolutely DG-exclusive manipulation. Rather, our findings indicate that PV-mediated inhibitory control within the hippocampal DG–CA3 circuit shapes the initial encoding of object–location representations and regulates the balance between sensitivity and consistency. The computational model formalizes this principle at the network level by testing how changes in inhibitory tone influence the formation, discrimination, and consistency of hippocampal-like memory representations. Thus, the anatomical spread of the viral manipulation limits the spatial specificity of the interpretation, but does not undermine the validity of the behavioral finding or the proposed sensitivity-consistency framework.

The different NOL-based paradigms used in this study were designed to probe complementary aspects of hippocampal-dependent memory processing under progressively increasing cognitive demands. Using the standard NOL task, we first established that PV⁺ interneuron modulation selectively affects memory encoding, without altering consolidation or retrieval processes. Building on this observation, we developed a sequential NOL paradigm to determine whether reduced PV-mediated inhibition altered the resolution of spatial discrimination. In this task, PV-Gi animals detected smaller spatial displacements than controls, indicating enhanced capacity to discriminate subtle spatial changes. We then asked whether this increased discrimination capacity carried a cost under conditions requiring the encoding and maintenance of multiple memory representations. To address this, we designed the multi-arena NOL task, which introduced increased contextual interference and representational demands through multiple object-location associations across distinct environments. Under these conditions, reduced PV-mediated inhibition impaired performance. Thus, while the standard and sequential NOL tasks indicate that reduced PV⁺ activity enhances discrimination when task demands are low to moderate, the multi-arena task reveals a performance cost when associative and interference demands increase. Together, these findings suggest that PV-dependent inhibitory dynamics contribute to balancing the sensitivity and consistency continuum.

A limitation of the multi-arena NOL task is that it cannot fully dissociate impaired contextual discrimination from impaired maintenance of multiple object–context–location associations. However, several observations argue against a generalized deficit in contextual discrimination. First, PV-Gi animals showed enhanced fine spatial discrimination in the low-load sequential NOL task, indicating that reduced inhibition does not impair discrimination per se. Second, in the multi-arena NOL task, PV-Gi animals did not fail uniformly across all arenas, but showed above-threshold discrimination in some contexts while retaining fewer contexts overall. This suggests that their ability to use contextual and spatial information was not abolished, but that performance was compromised when multiple object–context–location associations had to be maintained. Similarly, PV-Gq animals showed intermediate performance in the multi-arena NOL task rather than a generalized impairment, indicating that increased inhibition also does not abolish contextual discrimination. Together, these observations argue against a simple deficit in contextual discrimination. This interpretation is also consistent with previous work showing that manipulation of DG PV+ interneurons during contextual fear conditioning did not produce detectable deficits in long-term contextual memory discrimination, despite effective activation of the PV+ interneuron population [50].

Animals

All animal experiments were approved by the Animal Care and Use Committee of the Instituto de Neurociencias de Alicante (Alicante, Spain) and complied with the Spanish law (53/2013) and European regulations (EU directive 2010/63/EU).

All animals were two months old at the beginning of the experiments and were randomly assigned to different experimental groups (see below). We excluded animals showing low exploration behavior (see NOL task). No significant differences between sexes were found, and thus the data were pooled. Mice were bred in-house from the line 129P2-Pvalbtm1(cre)Arbr/J (Jackson Laboratories, RRID: IMSR_JAX:008069) and housed in groups (3−5), with 12−12 h light/dark cycle, lights on at 8:00, at room temperature (23 ± 2 ºC) and free access to food and water.

Viral injections

Electrophysiology

All electrophysiological protocols were recorded in the same animals before and after CNO injection (n = 6 PV-Gi, n = 6 PV-Gq, and n = 6 Sham). Mice were anaesthetized with urethane (1.4 g/kg, i.p.) and recording and stimulating electrodes were implanted following stereotaxic standard procedures. A bipolar stimulating electrode (WPI, ref. TM53CCNON) was placed in the PP with the following coordinates relative to bregma: −4.3 mm anteroposterior (AP), +2.5 mm mediolateral (ML), +1.4 mm dorsoventral [DV], and at an angle of −12º. The DV position was adjusted individually for each animal to optimize the evoked response. The recording probe had a single shank and 32 channels (50 μm contact spacing; NeuroNexus Technologies, Michigan, USA). It targeted the hippocampus (−2 AP, +1.5 ML, −2 DV), including CA1 and DG regions.

Electrophysiological recordings were made before and 90 min after CNO injection (1 mg/kg, i.p.). After post-CNO recording, we induced LTP using a standard theta burst protocol: six 400 Hz-trains of pulses delivered at a 200 ms interval, repeated six times with 20 s inter-train interval (Stäubli and colleagues, 1999). Evoked potentials were recorded 60 min after LTP induction again. We applied a stimulation intensity adjusted to 80% of the maximum PS.

Signals were filtered (0.1–3 k Hz), amplified and digitized (20–32 kHz acquisition rate), with data analyzed offline using Spike2 software (Cambridge Electronic, Cambridge, UK) or MatLab (The MathWorks, Natick, MA, USA).

To quantify the electrophysiological responses, we measured the EPSP as the maximum slope of the evoked potential in the molecular layer (for DG). The PS was measured as the amplitude of the fast negative potential recorded in the hilus of the DG. In the PV-Gi animals, in which CNO administration caused a large increase in PS amplitude, the EPSP was measured more distal from the granule cell soma layer (in the outer third of the molecular layer) to minimize the volume-conducted artifact of the PS, enduring more accurate EPSP slope measurements.

c-Fos experiments

To quantify the impact of PV modulation on c-Fos expression after encoding, 22 animals (8 PV-Sham, 8 PV-Gi, 6 PV-Gq) were allowed to freely explore a novel environment (a 50 x 50 x 30 cm arena containing spatial cues and two identical objects) and sacrificed 60 min later. Perfusion of the animals and tissue slicing were performed as described above. Eight to 10 serial slices, covering the hippocampus, PFC and NAc were processed for c-Fos immunostaining (primary antibody: guinea pig anti-cFos, 1:1000, ref. 226004, Synaptic Systems, Göttinger, Germany; secondary antibody: Alexa 488 donkey anti-guinea pig, 1:500, ref. 706-545-148, Jackson ImmunoResearch, Suffolk, United Kingdom). The slices were mounted and used in subsequent quantitative analysis.

Quantitative analysis of c-Fos positive nuclei was performed offline using 12-bit grayscale images acquired with a Leica DM4000 fluorescence microscope at 10x/0.25 dry objective. Images were analyzed using a Neurolucida software (MBF Bioscience, Williston, VT, USA). The ROIs were manually delineated following the Franklin and Paxinos mouse brain atlas [56]. Analysis was conducted using Icy Software [57] in a semi-automatic manner. The detection threshold for positive nuclei was set for each brain region, considering the average nuclei size and a signal-to-noise ratio greater than 23%, in accordance with the Rayleigh criterion for resolution and discrimination between two points. Animals with misplaced viral infections were excluded from all the analysis.

Histology

Animals were perfused with 50 ml of 37 ºC saline and 50 ml of 4% paraformaldehyde (PFA; BDH prolabo, VWR chemicals, Lovaina, Belgium). Brains were extracted and stored in 4% PFA at 4 ºC for at least 24 h. After fixation, brains were sectioned using a vibratome (VT 1000S, Leica, Wetzlar, Germany) into 50 μm slices. Validation of anatomical coordinates for recording and stimulating electrodes was done in DAPI stained slices. Viral infection efficiency and specificity were quantified in 6–8 hippocampal containing slices following standard immunohistochemistry protocols. For PV labeling, primary antibody (mouse anti-PV, 1:2000, ref. 235, Swant, Switzerland) and secondary antibody (Alexa 488 goat anti-mouse, 1:500, ref. A11029, Life Technologies; USA) were applied. Virus expression was further visualized using the mCherry reporter co-labeling.

Behavior

To assess the impact of PV modulation in spatial memory, we used a series of behavioral tasks. To quantify the animals’ exploratory behavior, we used a custom MATLAB-based code developed in the lab (RodEx, https://doi.org/10.20350/digitalCSIC/28955), which allowed animal tracking and automatic quantification of exploratory behavior.

Computational model

The computational model consisted of two modules [49]: one module with functionalities related to the hippocampus, such as learning and pattern separation, and another module with functionalities related to the prefrontal cortex, such as decision-making.

The hippocampal-related module was modeled as a neuronal layer with convergent projection from a “sensory” representation onto a memory space, with local competition between neurons [48,58]. Local competition was implemented using a k%-max winner-take-all mechanism [9], which simulates different levels of inhibition. In this mechanism, only the neurons whose activity falls within a k% of the maximum neuronal activity are activated. Higher values of k% reduce inhibition, increasing the probability of a neuron being active, while lower values increase inhibition.

The activity of 200 neurons simulates the “sensory” input pattern projected to the memory space. Each neuron is assigned a random activity value, drawn from a uniform distribution and normalized so that the total sum is 1. These neurons connect with a probability of 25%–500 neurons in the memory space. During learning, the synaptic weight between an active presynaptic neuron () and an active postsynaptic neuron () increases based on the activation values of both neurons and the learning rate ():

After each training step, the weights are normalized to their mean value. In the simulations, we used a learning rate of L_r = 0.05 to produce stable representations of learned patterns [49]. A postsynaptic neuron () in the memory space becomes active only if its input exceeds k% of the maximum input activity across all neurons in the layer. If this is the case, the neuron’s activity is set as the difference between its input activity and the k% from the maximum input activity, normalized by the maximum activity of all active neurons.

The hippocampal-related module was selected based on its previously proven validity to explain some of the dynamics taking place in the hippocampus, such as local competition or rate remapping [9,48]. The sum of the hippocampal neurons’ activity is projected onto the module associated with prefrontal cortex functionalities, where the decision is made.

The decision-making module is a mean-field approximation of a realistic network of integrate-and-fire neurons [59]. The rate-based model is a well-established approximation that allows the study of neuronal activity while reducing the computational expenses of modeling individual neurons. Moreover, it is supported by the fact that neurons are organized in neuronal populations with similar properties [60]. The model consists of two groups of excitatory neurons with recurrent connections (“explore object” and “explore arena” groups) that compete through mutual inhibition [61,62]. The activity of the excitatory groups follows the dynamics:

where 𝑦 represents the activity of each group, ω is the weight of the connections (ω+ recurrent; ω− inhibitory), represents the input from the hippocampal-related module to group “explore object”, c represents a constant input to the “explore arena” group, which mimics the activity from a learned (familiar) input pattern, 𝜎ξ represents the network fluctuations and f(.) is a sigmoidal function of the form:

where f_max defines the maximum activity, i.e., the activity at which the neuronal network saturates. The network fluctuations are simulated through a Gaussian noise with mean 0 and variance 1 (ξ), which simulates having a finite set of neurons, and a scale factor 𝜎 = 0.01 spikes·s⁻¹. In the simulations, we used: = 20 ms, 1, = 1, f_max = 40 spikes⋅s−1, = 4 spikes⋅s⁻¹, and k = 11 spikes⋅s⁻¹. Within the prefrontal-related module, the strength of the connections was kept constant during the simulations.

To simulate the DI during the testing sessions, the input to the “explore arena” group in the decision-making module was set constant, simulating the output of the hippocampal module for familiar patterns. The “explore arena” group received the output of the hippocampal module as the sum of the activity of its neurons. The input activities were scaled by a factor of 0.1. A decision towards exploring the object was made when the “explore object” group won the competition. The DI was calculated as the ratio between the number of decisions towards exploring the object when the input pattern was novel (not learned) and the number of decisions made towards exploring the object when the pattern was familiar (learned) or novel (total cases).

In the simulations, we used 20 “sensory” input patterns to train the hippocampal-related module, except when investigating memory stability, where the number of patterns varied. During learning, each pattern was presented as input for 50 trials and was subsequently referred to as a familiar pattern. After learning, the same or a specifically modified familiar pattern was presented for 30 runs at the input to test the model. These modified patterns were generated by introducing different levels of noise to the original input pattern, resulting in patterns with different degrees of similarity to the familiar one. Similarity between sensory representations was quantified as the correlation between the activity vector of familiar and modified input patterns. For each level of input similarity and inhibition (competition) levels, the entire procedure was repeated 10 times. The similarity of memory representations was then quantified as the correlation between the output evoked by a familiar pattern and the outputs evoked by gradually modified versions of that pattern.

To simulate the sequential NOL task, the model first learned 20 input patterns over 50 runs, as previously. The model was then tested over 30 runs, as described above, with these learned familiar patterns (familiar condition) and with noisy variants exhibiting a correlation of 0.8 with one of the familiar patterns. These noisy patterns were subsequently used for a second learning phase (10 runs), after which the model was again tested over 30 runs using newly generated noisy patterns with a correlation of 0.6 between familiar and those new patterns. As above, the entire procedure was repeated 10 times.

To evaluate the model’s robustness to noise, we varied the strength of the connections between the input and memory space within the hippocampal-related module. Each connection weight was modified by adding a pseudorandom value between 0 and 1, scaled by a factor ranging from 0 to 5 in steps of 1.

Memory interference was assessed by increasing the number of patterns from 50 to 1,000, in increments of 50, for each level of competition.

Statistical analysis

The statistical analysis was done using GraphPad Prism 7 software (GraphPad Software, La Jolla, CA, USA) or SPSS v20 software (IBM, New York, USA). After an exploratory analysis for the presence of outlier values, we checked the skewness and kurtosis of the data before testing their statistical distribution. Parametric-test requirements, including normality (D’Agostino-Pearson test and Shapiro-Wilk test) and homoscedasticity (F of Levene test) were tested. All the data fulfilled parametric criteria, unless otherwise specified. In most analyses we applied two-way ANOVA, with a group factor with 3 levels (Sham, PV- Gi and PV-Gq), and a time factor with 2 levels (before and after CNO injection). In case of c-Fos analysis, we applied one-way ANOVA with group as a factor with the same 3 levels as above. We applied a Sidak post hoc analysis for multiple comparisons with adjusted alpha. Effect size was calculated with partial eta square (ηp²) as SSeffect / (SSeffect +SSerror), and the power of the effect (1-β) using GPower software (University of Düsseldorf, Düsseldorf, Germany). Effect size is considered large if ηp² > 0.13. If, by any reason, partial eta square could not be used, then we express r² indicator, being similar to eta, but requiring >0.30 values to indicate large effects. The power of the effect (1-β) is taken as an invert indicator of committing a type II error. The results of the statistical analysis are detailed in Supp. Tables numbered according to the number of the Figures to which they correspond.

Experimental data were plotted using GraphPad Prism 7 software (GraphPad Software) and figures were composed using Adobe Illustrator software (Adobe Systems Incorporated, San José, California, USA).

Significance statement

Memory systems must balance sensitivity, to detect subtle changes, and consistency, to preserve stable, interference-resistant representations. How the brain dynamically regulates this trade-off remains unclear. Here, we show that inhibition in the dentate gyrus regulates this balance. Combining cell-type-specific pharmacogenetics, behavioral assays, and computational modeling, we found that reducing inhibition during encoding enhanced subsequent discrimination of similar experiences but also increased vulnerability to interference. In contrast, increased inhibition stabilized memory representations but reduced discrimination. These findings establish inhibition as a key modulator of hippocampal memory processing, enabling the system to prioritize either flexible updating or robust retention, with implications for understanding cognitive function and dysfunction.

S1 Fig. Expression of DREADDs in the DG of PV-cre animals.

(a) Representative pictures showing the reporter of the virus infection (mCherry, red), immunolabelled against PV protein (green) and counterstained with DAPI (cell nuclei, blue). Scale bar: 100 μm. (b) Zoom in of an infected PV⁺ cell (white square in A). Scale bar: 10 μm. (c) Efficiency of DREADDs expression in the DG expressed as the % of infected PV-cells in PV-Gi (orange) and PV-Gq (blue) animals. Data show mean ± SEM. The data underlying this Figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3003956.s001

(TIF)

S2 Fig. Modulation of PV-cell activity selectively affects trisynaptic but not disynaptic responses recorded in CA1.

(a) Schematic representation of the disynaptic (PP – CA3 – CA1; magenta) and trisynaptic (EC – DG – CA3 – CA1; black) hippocampal circuits. (b) Evoked potentials recorded in CA1 stratum radiatum in response to perforant pathway (PP) stimulation. Magenta and black arrows point to di- and tri-synaptic EPSPs, respectively. (c) Comparison of disynaptic EPSP recorded after PP stimulation (average of 6 single pulses at 800 μA) before and after CNO administration. (d) Comparison of trisynaptic EPSP evoked with the same stimulation as in c. Data show mean ± SEM. All statistical values are detailed in S3 Table. The data underlying this Figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3003956.s002

(TIF)

S3 Fig. Quantification of activated c-Fos+ cells.

(a) Temporal sequence of the experimental protocol for c-Fos+ cells quantification. After 90 min post-CNO injection, animals were either allowed to explore a novel arena with 2 objects (encoding phase of a NOL task) or maintained in their home cage (HC). (b) Quantification of modulated PV⁺ cell activity, expressed as the percentage of PV⁺ cells in the DG that are double stained for c-Fos. (c) Example of activated c-Fos+ cells in DG in HC and after NOL encoding phase (context exploration, CE). Scale bar: 100 μm (d) Number of activated c-Fos+ cells in the DG during the NOL encoding phase versus HC. Scale bar: 100 μ. All statistical values are detailed in S3 Table. The data underlying this Figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3003956.s003

(TIF)

S4 Fig. Motor activity and anxiety.

(a) Diagram of the Novel Object Location (NOL) protocol followed. (b) Detailed diagram of the NOL arena. Example of one mouse tracking within a test session (bottom). (c) For evaluating locomotor activity, with and without PV modulation, the traveled distance in each session was analyzed for the three groups of mice. (d), Diagram of the Elevated Plus Maze used for evaluating anxiety (top). Example of one trial (bottom). (e) Anxiety levels as measured by time spent, distance traveled, resting time and entries per zone. All statistical values are detailed in Supp. S3 Table. The data underlying this Figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3003956.s004

(TIF)