The study, published in PLOS Biology, was led by Santiago Canals, head of the Plasticity of Brain Networks laboratory, together with Encarni Marcos, who heads the Neural Mechanisms of Behaviour research line at the same institute. Their findings show that the level of inhibitory tone in the dentate gyrus, a gateway structure within the hippocampus, determines whether the memory system prioritises sensitivity to small differences between experiences or consistency in the face of change. Crucially, the researchers discovered that this is not a fixed setting but a dynamic dial, one whose optimal position depends on how much information the brain is being asked to remember.

Memory, as Canals emphasises, is not a static repository of information but a living and adaptive process. For memory to be useful, the brain cannot simply store a perfect copy of every experience. If every minor variation in the environment generated an entirely new memory trace, we would be unable to recognise our own homes after the slightest rearrangement of furniture. At the same time, the system must remain sensitive enough to distinguish between two experiences when the differences genuinely matter, such as recognising that a familiar route has been blocked or that a normally friendly face is showing signs of anger. The dentate gyrus, the researchers argue, is one of the key places where this compromise is negotiated.

The dentate gyrus occupies a special position in hippocampal circuitry. It receives most of the information flowing into the hippocampus from the surrounding cortex and performs a transformation that neuroscientists call pattern separation, the process by which similar inputs are made more distinct from one another before they are passed deeper into the memory system. Inhibitory interneurons, the cells that suppress the activity of their neighbours, are central to this computation. By experimentally modifying the activity of these inhibitory cells in mice during memory encoding, the window of time in which information is processed to form a new memory, the team was able to shift the balance between sensitivity and consistency in a controlled way.

The behavioural assays the researchers used exploited a natural preference of mice for novelty. When a mouse encounters something it has not seen before, it spends more time exploring it. By presenting animals with objects whose positions had been shifted by varying amounts, the team could infer from exploratory behaviour whether the animals remembered the original arrangement and how finely they could discriminate changes to it. The results were striking. When inhibition was reduced below normal levels, mice became exquisitely sensitive to change: they detected displacements of objects so small that control animals ignored them entirely. Their behaviour, as Marcos describes it, was more consistent with better memory retrieval and more detailed recall. Conversely, when inhibitory activity was increased, the animals became less sensitive to small differences, and their memory representations proved more resistant to change.

At first glance, these results might suggest that lowering inhibition is simply beneficial, producing sharper, more detailed memories. But the team suspected the story was more complicated, and to test that suspicion they turned to computational modelling. The model offers a simplified mathematical representation of how memory representations are formed and retrieved, allowing the researchers to systematically vary the level of inhibition and observe the consequences for two competing performance measures: the ability to distinguish between similar experiences, and the ability to maintain consistent representations across changing conditions. The simulations revealed a fundamental constraint. Lower inhibition improves discrimination, but it also makes the stored representations increasingly vulnerable to interference, the corruption of one memory by another, and this vulnerability grows as the number of items to be remembered increases.

The model generated a specific and testable prediction: there is no single optimal level of inhibition that works in all situations. When memory load is low, reduced inhibitory activity can favour fine discrimination between similar experiences. But as the amount of information to be remembered grows, higher inhibitory activity becomes advantageous because it helps maintain representations that are sufficiently consistent to avoid confusion between memories. In other words, the ideal operating point of the dentate gyrus shifts depending on the cognitive demands of the moment, a prediction that departs from the simpler view of inhibition as a fixed filter.

To test this prediction in living animals, the researchers designed a more demanding task in which mice had to learn associations between objects, locations and different environments, thereby increasing the memory load relative to the simpler object-placement assays. The behavioural results matched the model’s forecasts. Under conditions of higher memory load, the advantages of reduced inhibition diminished and reversed, exactly as the computational framework predicted. This convergence between simulation and experiment is one of the study’s most significant achievements. By combining experimental and computational tools, as Canals explains, the team identified a balance point in memory formation that helps explain when this inherently flexible system updates with new information and when it instead preserves greater consistency.

An important detail of the findings concerns timing. The effects of manipulating inhibition were observed specifically during encoding, the phase in which the initial representation of an experience is constructed. The mechanism identified therefore shapes how memories are built in the first place, rather than how they are later stabilised through consolidation or retrieved from storage. This distinction matters for how neuroscientists think about the dentate gyrus: rather than being a passive relay, it appears to actively set the operating characteristics of new memories at the moment of their creation, determining in advance whether a given representation will lean toward fine-grained discrimination or robust stability.

Taken together, the results suggest that inhibitory circuits in the dentate gyrus place the hippocampus in different functional states. In a high-sensitivity state, the system readily incorporates changes and discriminates between similar experiences, at the cost of greater vulnerability to interference. In a high-consistency state, representations are protected from interference but small differences may go unnoticed. Marcos interprets the findings as evidence for a mechanism that dynamically adjusts the mode of operation by using inhibition, effectively retuning the memory system as task demands change. The researchers are careful to note, however, that the study has not yet identified which signals control this adjustment under natural conditions. Determining which regulatory mechanisms modulate inhibitory activity, and when in the course of natural behaviour the brain shifts between these states, will be the subject of future research. The work was funded by the Spanish State Research Agency, the Ministry of Science, Innovation and Universities, the Severo Ochoa Programme for Centres of Excellence, the European Union’s Next Generation funds, and the Generalitat Valenciana. Beyond its immediate implications for basic neuroscience, the findings may eventually inform our understanding of disorders in which this balance goes awry, from conditions of excessive memory generalisation to those marked by disruptive interference between memories.

Subject of Research: The role of dentate gyrus inhibitory interneurons in balancing memory flexibility and stability

Article Title: Study by the Institute for Neurosciences CSIC-UMH reveals how the brain balances memory flexibility and stability

Article References: Study by the Institute for Neurosciences CSIC-UMH reveals how the brain balances memory flexibility and stability. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: neuroscience, hippocampus, dentate gyrus, memory, inhibitory interneurons, pattern separation, memory encoding, computational model, PLOS Biology, CSIC, memory load, interference

The research team set out to test whether the famous separation principle holds equally along the full length of the hippocampus, a seahorse-shaped structure deep in the brain that is essential for storing memories of places and events. Neuroscientists have long divided the hippocampus into a dorsal pole, which in rodents and humans is heavily involved in fine-grained spatial and cognitive processing, and a ventral pole, which is more intimately connected with emotion, stress, and fear circuitry. The two regions differ in their connectivity, their gene expression profiles, and even the electrical properties of their principal neurons. Yet most theories of memory encoding were built on data from the dorsal end, leaving a crucial question unanswered: does the emotional half of the hippocampus follow the same coding rules as the cognitive half?

To find out, the researchers turned to a state-of-the-art combination of behavioral design and miniaturized calcium imaging. They implanted tiny microscopes onto the heads of mice and tracked the activity of hundreds of individual neurons in CA1 — the hippocampus’s main output layer — in both the dorsal and ventral regions simultaneously. Each mouse learned to associate one experimental context, designated context A, with mild foot shocks, while a second, visually distinct context B remained neutral and safe. The animals responded exactly as expected, freezing with fear in context A and exploring calmly in context B. But the neural story underneath that behavior proved far more surprising than any textbook prediction.

When the researchers compared the population activity patterns — the combined firing fingerprints of all recorded neurons — before and after fear conditioning, they found that both dorsal and ventral CA1 changed their representations of the shock-paired environment. Fear learning literally rewrote the spatial map. The critical difference was magnitude. In ventral CA1, the representational shift was substantially larger than in dorsal CA1, and the size of the shift scaled with how strongly each mouse froze: the more the animal feared the context, the more dramatically its ventral map had reorganized. The dorsal maps also changed, but they retained their hallmark quality of staying crisp and stable across repeated exposures.

The real challenge to the canonical theory came during discrimination testing, when mice were shuttled between the threatening and neutral contexts. In dorsal CA1, the two context representations remained clearly distinct, exactly as pattern separation theory demands. In ventral CA1, however, the opposite occurred: the representations of the dangerous and the safe environment became more similar to each other. Rather than pushing the two memories apart to prevent interference, fear learning in the ventral hippocampus pulled them together, creating a zone of overlap between the neural code for threat and the neural code for safety.

The authors interpret this overlap through the mathematical language of attractor dynamics — a framework borrowed from theoretical neuroscience in which stable patterns of network activity behave like valleys in an energy landscape. A neural representation, on this view, is a basin into which activity naturally settles, and the deeper and more sharply separated the basins, the more energy it takes to jump between them. Before fear conditioning, both dorsal and ventral CA1 hold well-separated context representations, and switching between them requires crossing a substantial energy barrier. After conditioning, dorsal CA1’s landscape is essentially unchanged. But ventral CA1’s landscape flattens: the basin corresponding to the threatening context widens and migrates closer to the neutral one, opening more entry points and lowering the energy required to fall into the fear state.

The consequence of this shallower landscape is speed, and the study demonstrates it directly. By analyzing how quickly each region reinstated the threatening context representation when mice transitioned back into context A, the researchers found that ventral CA1 expressed the fear-context code faster and more strongly than dorsal CA1. Even more striking, the relative dominance of the threatening versus neutral representation in ventral CA1 predicted how much each mouse actually froze — a direct link between the geometry of neural codes and the expression of fear behavior. When mice moved into the neutral context, the same logic ran in reverse: ventral CA1 showed the most rapid and pronounced suppression of the threatening representation among the regions examined, consistent with overlapping codes being quick to both engage and disengage.

It is tempting to see this as a computational trade-off, and the authors make that trade-off explicit. Pattern separation and pattern completion have always been two sides of the same coin in hippocampal theory: separating representations guards against interference, while allowing representations to blend permits a memory to be triggered by partial or ambiguous cues. The ventral hippocampus, this work suggests, tilts the balance heavily toward completion. For an animal whose survival depends on detecting danger quickly, the cost of occasionally mistaking a safe context for a dangerous one may be far smaller than the cost of failing to recognize a lethal threat in time. Rapid retrieval of fear memories, even at the price of some discriminative precision, may be an evolutionarily rational bargain.

The findings also help resolve a long-standing tension in the literature on hippocampal fear. Earlier work showed that ventral CA1 contains ensembles whose correlated activity retrieves contextual fear memories, that ventral hippocampal projections to the prefrontal cortex and amygdala regulate anxiety and avoidance, and that inactivating the ventral hippocampus alters fear expression and extinction. What remained unclear was how these emotional functions coexisted with the region’s evident role in spatial coding. The new results propose a unifying principle: the ventral hippocampus encodes space, but it warps its spatial codes in the service of valence, reshaping them so that emotionally significant environments become easier to summon from memory.

The implications extend beyond basic science. Overlapping or overly generalized context representations are a suspected hallmark of maladaptive fear in conditions such as post-traumatic stress disorder, where innocuous environments come to trigger full-blown threat responses. If the ventral hippocampal mechanism described here generalizes to humans, therapies aimed at restoring separability between threat and safety codes — rather than merely dampening fear output — could offer a more targeted route to treatment. For now, the study delivers its most memorable lesson in the cleanest terms: sometimes the brain remembers better not by keeping memories apart, but by letting danger and safety share the same neural ground, ready to tip into fear the instant the world turns threatening.

Subject of Research: How fear learning reshapes spatial context representations in the ventral versus dorsal hippocampus to enable rapid fear memory retrieval in mice

Article Title: Overlapping representations in the ventral hippocampus support rapid fear memory retrieval

Article References: Rozeske, R. R., Runtz, L., Lee, J. Q., Keinath, A. T., Sossin, A., & Brandon, M. P. (2026). Overlapping representations in the ventral hippocampus support rapid fear memory retrieval. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02435-5

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02435-5

Keywords: hippocampus, fear conditioning, memory retrieval, place cells, calcium imaging, attractor dynamics, pattern separation, ventral CA1, dorsal CA1, neuroscience, spatial memory, contextual fear

Memory formation and retrieval depend heavily on the brain’s remarkable ability to represent sequences of events, a function critical for everything from language processing to planning complex actions. Previous models have suggested that neurons engaged in sequence encoding showcase their activity in distinct phases of underlying rhythmic brain activity, such as theta or gamma oscillations. These phase relationships were posited to create a neural “timestamp,” marking the position of an event within a string of occurrences. This paradigm has heavily influenced experimental design and computational frameworks in neuroscience and artificial intelligence alike.

The research team, headed by Liebe, Niediek, and Pals, employed a multi-modal approach combining intracranial electrophysiological recordings from human subjects with rigorous computational modeling using recurrent neural networks (RNNs). These human data were acquired during tasks requiring participants to memorize and recall sequences of visual stimuli, enabling unprecedented insight into the neural dynamics underpinning sequence memory. Complementary RNN simulations mirrored these cognitive demands, allowing comparisons between biological and artificial systems under matched conditions.

Their analysis revealed a surprising dissociation: neurons’ firing phase relative to ongoing oscillations did not reliably reflect the serial position of elements within a memorized sequence. Instead, firing phase appeared more tightly associated with other coding dimensions, such as stimulus identity or the detection of salient events within the sequence. Interestingly, recurrent neural networks trained to emulate human sequence memory exhibited similar patterns, with phase-based activity failing to encode temporal order distinctly but rather highlighting identity and saliency features.

This revelation calls into question the conventional interpretation of phase coding as a simple temporal marker in episodic memory circuits. It suggests that brain oscillations might orchestrate sequence memory formation through a multiplexed coding scheme, where timing within oscillatory cycles contributes to richer representational codes beyond linear order. Such multiplexing could provide neural circuits with a flexible framework to interact with diverse cognitive variables, balancing temporal placement, item properties, and context-dependent salience.

The implications extend beyond basic neuroscience, impacting how artificial intelligence systems model human cognition. Recurrent neural networks are pivotal architectures in machine learning, particularly for sequential data processing. Understanding their phase-like activity patterns and their limitations in temporally encoding sequences offers a blueprint for novel mechanistic models. These models would shift away from simplistic temporal tagging toward integrative coding strategies, potentially enhancing memory recall fidelity and sequence prediction in advanced AI.

At the cellular level, intrinsic neuronal properties and synaptic dynamics likely contribute to the observed complex phase relationships. The study highlights how oscillatory timing interacts with synaptic inputs and intrinsic excitability to shape firing patterns that do not strictly follow temporal sequence boundaries. This complexity underscores the need for refined biophysical models that capture how neural populations encode multidimensional information in a temporally dynamic fashion.

Moreover, the interdisciplinary approach blending human neurophysiology with AI modeling exemplifies a new paradigm in cognitive neuroscience. Such integrative methodologies help reconcile discrepancies between animal models, human recordings, and computational theories, offering a holistic understanding of memory mechanisms. The parallel observation of phase coding properties in human neurons and artificial neurons strengthens the validity of these findings and challenges existing theoretical frameworks.

Overall, this study reframes the role of neuronal firing phase in encoding episodic sequences, emphasizing a departure from a unidimensional temporal code toward a participatory coding scheme involving identity and contextual salience. It provokes a reconsideration of how temporal information is embedded into neural circuits and invites further inquiry into the oscillatory dynamics facilitating complex memory tasks.

The authors underscore that while firing phase does not strictly encode temporal order, oscillatory activity still plays a crucial role in orchestrating neural ensemble dynamics, possibly gating information flow or segmenting activity to optimize memory processes. Future research will be needed to elucidate precisely how different oscillation bands and their interactions contribute to multiplexed coding schemes supporting flexible, accurate sequence memory.

This paradigm shift aligns with emerging evidence from other cognitive domains, where timing relative to oscillations intertwines with other neural coding dimensions like amplitude modulation, synchrony, and spike timing plasticity. Together, these mechanisms likely converge to enable the brain’s astounding capacity to represent and manipulate complex event sequences in real time.

The findings pose exciting challenges and opportunities for designing biologically inspired AI systems capable of richer temporal representations. Embedding multiplexed phase coding principles into recurrent architectures may yield breakthroughs in natural language processing, robotics, and memory augmentation technologies.

In closing, this research marks a significant advancement toward unraveling the neural code underlying human sequence memory. By demonstrating that phase of firing does not straightforwardly reflect temporal order, it invites the scientific community to explore new avenues in oscillatory coding and memory research that bridge biology and artificial intelligence.

As our understanding deepens, it becomes increasingly evident that brain oscillations do not merely tick off moments on a mental clock but rather sculpt a dynamic and multi-layered tapestry of neural information, threading together identities, times, and contexts into the cohesive fabric of human experience.

Subject of Research: Sequence memory and neural coding mechanisms in humans and recurrent neural networks.

Article Title: Phase of firing does not reflect temporal order in sequence memory of humans and recurrent neural networks.

Article References:
Liebe, S., Niediek, J., Pals, M. et al. Phase of firing does not reflect temporal order in sequence memory of humans and recurrent neural networks. Nat Neurosci 28, 873–882 (2025). https://doi.org/10.1038/s41593-025-01893-7

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s41593-025-01893-7