Announcer The following program features simulated voices generated for educational and philosophical exploration.

Adam Ramirez Good evening. I'm Adam Ramirez.

Jennifer Brooks And I'm Jennifer Brooks. Welcome to Simulectics Radio.

Adam Ramirez Tonight we're examining working memory—the cognitive capacity to maintain information over seconds when it's no longer perceptually available. For decades, the dominant model held that working memory depends on persistent neural activity, where neurons continue firing throughout the delay period to actively maintain representations. Recent work challenges this view, proposing that information might instead be stored in short-term synaptic changes that can be read out when needed without sustained firing. The question is whether working memory requires continuous neural activity or whether transient synaptic modifications provide sufficient substrate for temporary storage.

Jennifer Brooks The persistent activity model emerged from single-neuron recordings in prefrontal cortex during delayed-response tasks. Neurons fire throughout the delay period between stimulus and response, with firing rate encoding specific stimulus features like location or identity. This sustained activity seemed like a neural correlate of active maintenance—the brain literally keeping information 'online' through continuous spiking. However, several observations complicate this picture. First, persistent activity is metabolically expensive. Second, activity patterns during delays are often dynamic rather than stable. Third, brief perturbations don't always disrupt memory. These findings have motivated alternative models where information resides in synaptic weights that can be rapidly modified and then read out by transient activity bursts.

Adam Ramirez To explore whether working memory depends on persistent activity or synaptic mechanisms, we're joined by Dr. Christos Constantinidis, a neuroscientist at Vanderbilt University whose research focuses on neural mechanisms of working memory and attention in primate cortex. His work combines electrophysiology, computational modeling, and behavioral testing to understand how information is maintained over short delays. Dr. Constantinidis, welcome.

Dr. Christos Constantinidis Thank you. The debate between activity-based and synapse-based working memory touches fundamental questions about neural coding and information storage.

Jennifer Brooks Let's start with the classic evidence. What do single-neuron recordings reveal about prefrontal cortex activity during working memory tasks?

Dr. Christos Constantinidis When you record from prefrontal cortex while monkeys perform delayed-response tasks—where they see a stimulus, wait through a delay, then make a response—many neurons show elevated firing during the delay period. The firing rate often correlates with specific stimulus features. For instance, if the monkey needs to remember a location, neurons tuned to that location maintain higher firing throughout the delay. This persistent activity survives across the seconds-long interval between stimulus offset and response cue, suggesting it actively maintains the memory trace. The strength and duration of persistent activity predicts behavioral performance—stronger, more stable activity corresponds to better memory accuracy. This evidence strongly supports the idea that sustained firing implements working memory.

Adam Ramirez But if persistent activity is the mechanism, how does the circuit maintain stable firing for seconds? What prevents activity from either decaying or running away?

Dr. Christos Constantinidis This is where network models become important. The standard account involves recurrent excitation—neurons excite each other through positive feedback loops, creating a self-sustaining activity pattern. However, you need inhibition to balance excitation and prevent runaway firing. Models incorporate tuned excitation and broader inhibition to create stable attractor states where activity can persist at specific levels corresponding to different memory contents. These attractors need to be stable enough to maintain activity but flexible enough to update when new information arrives. The circuit must also handle noise, which can destabilize activity bumps. Theoretical work shows this is achievable with appropriate connectivity patterns, but experimental verification of the predicted connectivity has been challenging.

Jennifer Brooks What about the metabolic cost argument? Sustaining elevated firing for seconds across populations seems energetically expensive.

Dr. Christos Constantinidis This is a valid concern. Persistent activity does consume energy—maintaining elevated spiking and the associated ion pumping required to restore resting potentials is metabolically demanding. However, several points are worth considering. First, not all neurons need to fire persistently. Sparse coding models propose that a small fraction of the population carries the memory signal, reducing total energy expenditure. Second, the firing rates during delays are often modest—typically tens of spikes per second rather than hundreds—which is less expensive than maximal firing. Third, working memory is a temporary state. We're not maintaining information for minutes or hours, just seconds, so the total energy cost per trial might be manageable. That said, if synaptic mechanisms could achieve the same function with lower energy cost, that would be evolutionarily advantageous.

Adam Ramirez Let's turn to the synaptic storage alternative. What's the proposed mechanism, and how would it work?

Dr. Christos Constantinidis The core idea is that working memory relies on activity-silent mechanisms—short-term changes in synaptic efficacy that don't require sustained firing. During the stimulus period, neural activity triggers rapid synaptic modifications, perhaps through calcium-dependent processes or changes in receptor availability. These modifications persist after stimulus-evoked activity subsides, creating a latent memory trace. When a retrieval cue arrives, transient activity reads out the synaptic state, reactivating the memory representation. Between stimulus and retrieval, there's little or no elevated firing, making the memory 'silent' in terms of spiking activity but present in synaptic weights. This proposal is motivated partly by findings that brief perturbations during delays don't always erase memories, suggesting the information isn't solely in active firing patterns.

Jennifer Brooks What experimental evidence supports activity-silent mechanisms?

Dr. Christos Constantinidis Several types of evidence have emerged. First, some experiments using transcranial magnetic stimulation in humans show that disrupting cortical activity during the delay doesn't completely abolish memory, and subsequent reactivating pulses can restore memory performance, suggesting a latent trace survived the perturbation. Second, studies in rodents have found that optogenetically silencing neural populations during delays impairs but doesn't eliminate memory, with performance recovering when silencing ends. Third, computational models demonstrate that synaptic mechanisms like short-term facilitation can maintain information over seconds through changes in release probability without sustained presynaptic activity. However, interpreting these experiments is complex because silencing techniques may be incomplete or may indirectly affect synaptic state.

Adam Ramirez How do we reconcile the robust persistent activity recordings with the synaptic storage hypothesis? If neurons are firing during delays, isn't that evidence against activity-silent storage?

Dr. Christos Constantinidis Not necessarily. The two mechanisms aren't mutually exclusive. One possibility is that both contribute, with their relative importance varying by brain region, task demands, or memory load. For instance, prefrontal cortex might rely more on persistent activity for tasks requiring active manipulation, while parietal cortex might use more synaptic storage for simple maintenance. Another possibility is that what appears as sustained activity in population averages actually consists of sparse, stochastic bursts from individual neurons. If synaptic changes bias the probability of these bursts, the population shows elevated firing even though individual neurons aren't continuously active. The memory would then reside partly in synaptic weights and partly in the ongoing stochastic reactivation. Distinguishing these scenarios requires analyzing fine-scale temporal structure in spiking patterns.

Jennifer Brooks What about the dynamics of delay activity? Some recordings show that activity patterns aren't stable but change continuously during delays.

Dr. Christos Constantinidis This is an important observation. Early persistent activity models assumed static, stable firing throughout the delay. But more recent work reveals that population activity often evolves dynamically—neurons change their firing patterns in systematic ways even when the memory content is unchanging. These dynamics might reflect sequential processing stages, with different neural ensembles active at different delay phases. One interpretation is that dynamic patterns provide temporal context information in addition to memory content. Another possibility is that dynamics reflect the network exploring a low-dimensional manifold where memory is encoded in the trajectory through state space rather than a fixed firing pattern. These dynamic codes could still implement working memory through persistent activity, but the coding scheme is more complex than initially thought.

Adam Ramirez Does working memory capacity relate to these mechanisms? If there's a limit to how much we can hold, does that favor activity-based or synapse-based models?

Dr. Christos Constantinidis Working memory capacity limits are well-established—humans typically maintain about three to four items robustly. The mechanistic basis of this limit differs between models. In persistent activity models, capacity might be limited by the number of stable attractor states the network can support simultaneously. Each memory item requires a population of neurons maintaining elevated activity, and interference between items or limited neural resources constrains capacity. In synaptic models, capacity might be limited by the number of independent synaptic changes that can be induced and maintained without interference. Both mechanisms face similar constraints from representational overlap and noise. However, synaptic models might predict different patterns of interference between sequentially encoded items, which could be tested experimentally.

Jennifer Brooks How do perturbation experiments distinguish between mechanisms? If you disrupt activity during delays, what should happen under each model?

Dr. Christos Constantinidis This is the key experimental test. Pure persistent activity models predict that disrupting spiking should erase the memory—if information exists only in ongoing firing, silencing neurons should eliminate it. Pure synaptic models predict that brief perturbations should have minimal lasting effects because information resides in synaptic weights, which aren't immediately reset by activity cessation. However, hybrid models complicate predictions. If synaptic changes require periodic reactivation to be maintained, transient silencing might degrade but not eliminate memory. If activity patterns can be restored from synaptic weights after perturbation ends, memories might recover. The challenge is achieving sufficiently complete and selective perturbations while monitoring both spiking and synaptic state, which is technically difficult.

Adam Ramirez Are there computational advantages to one mechanism over the other? Would one enable operations the other couldn't perform?

Dr. Christos Constantinidis Persistent activity might offer faster read-out because information is already in the spiking domain, readily accessible for downstream processing or decision-making. It also allows continuous monitoring and updating—active representations can be flexibly combined with new inputs. Synaptic storage might be more energy-efficient and could enable larger capacity by distributing information across more synapses than could be simultaneously active. It might also be more robust to transient perturbations. However, retrieving synaptic memories requires specific reactivation signals, which could add latency or complexity. For computations requiring rapid, flexible manipulation of maintained information—like mental arithmetic or planning—persistent activity seems advantageous. For simple storage without manipulation, synaptic mechanisms might suffice.

Jennifer Brooks How do different brain regions contribute? Is working memory anatomically localized, or is it a distributed process?

Dr. Christos Constantinidis Working memory involves distributed networks. Prefrontal cortex, particularly dorsolateral prefrontal cortex, has been emphasized because of robust persistent activity during delays. But parietal cortex, especially lateral intraparietal area, also shows delay activity. Sensory cortices can maintain stimulus-specific information. Subcortical structures like basal ganglia and thalamus interact with cortical networks. One model proposes functional specialization—prefrontal cortex maintains task rules and goals, parietal cortex maintains spatial information, temporal cortex maintains object information. Different regions might employ different mechanisms. Evidence suggests prefrontal cortex relies more on persistent activity, while early sensory cortex might use more synaptic storage. This heterogeneity makes sense given different computational demands across regions.

Adam Ramirez What role does attention play? Is working memory maintenance the same as attending to internally represented information?

Dr. Christos Constantinidis There's substantial overlap between working memory and attention mechanisms. Both involve selective enhancement of task-relevant representations. During working memory delays, attention is presumably directed to the memory trace to maintain it against interference. Electrophysiological signatures of attention—like increased firing rates and decreased trial-to-trial variability—are also observed for working memory representations. Some theories propose working memory is simply sustained attention applied to internal representations rather than external stimuli. However, attention can shift between different memory items or to external events while memories persist, suggesting some independence. The relationship likely depends on mechanism—persistent activity might require continuous attentional resources to sustain firing, while synaptic storage might allow attention to disengage temporarily.

Jennifer Brooks Are there pathological conditions that illuminate the mechanisms? What happens in disorders affecting working memory?

Dr. Christos Constantinidis Schizophrenia involves prominent working memory deficits, and there's evidence of reduced delay-period activity in prefrontal cortex. This has been interpreted as weakened persistent activity due to altered excitatory-inhibitory balance, possibly from NMDA receptor dysfunction. The synaptic storage hypothesis would predict different pathophysiology—perhaps impaired synaptic plasticity mechanisms. Aging also impairs working memory, with reduced persistent activity amplitude and stability in older subjects. Pharmacological manipulations affecting different neurotransmitter systems have varying effects on delay activity and performance, potentially informing mechanisms. However, pathological changes are usually complex, affecting multiple systems, making it difficult to isolate specific mechanisms.

Adam Ramirez Looking forward, what experiments would definitively test whether working memory requires persistent activity or can be implemented synaptically?

Dr. Christos Constantinidis We need experiments that independently manipulate spiking activity and synaptic state during delays. Optogenetic tools that can reversibly silence neurons while ideally preserving or monitoring synaptic changes would be valuable. We could silence activity, then test whether memories can be retrieved by subsequent reactivation. Complementary approaches might use pharmacological agents that affect synaptic plasticity mechanisms without directly silencing neurons. Another direction involves high-resolution imaging of synaptic markers during working memory tasks to see whether rapid synaptic changes correlate with memory maintenance independent of sustained firing. Computational modeling that makes divergent predictions between mechanisms for specific experimental manipulations would help target experiments. Ultimately, we may find working memory relies on both mechanisms in different contexts.

Jennifer Brooks Dr. Constantinidis, thank you for clarifying what we know about working memory mechanisms and what remains experimentally open.

Dr. Christos Constantinidis Thank you. Working memory sits at the intersection of perception, attention, and action, and understanding its mechanisms is fundamental to understanding cognition.

Adam Ramirez That's our program. Until tomorrow, stay critical.

Jennifer Brooks And keep questioning. Good night.