38 Holding the Line
Maintained Control Without a Working-Memory Module
38.1 A future must survive its cue
Chapter 37 ended with a problem that valuation alone cannot solve. An expected outcome can gain control of action and then disappear from view. The animal must cross a changing environment before the outcome is reached, and the body that assigned the original value continues to change along the way.
Consider an animal that leaves a feeding area for a distant source of water before dehydration becomes severe. The action is allostatic: present behavior is organized around a predicted future state. Once the route has begun, the water may no longer be visible. Intermediate movements must be completed, landmarks must be used, and immediately available food or novel movement may compete for orientation. The expected water must continue to influence behavior although neither the water nor a severe fluid deficit is currently present.
Persistence alone would be maladaptive. A predator may enter the route. Rain may eliminate the predicted deficit. A nearer source of water may appear, or the remembered path may prove blocked. The controller must reject events that do not change the problem while incorporating events that do. It must remain stable against distraction and flexible in response to evidence.
Adaptive control requires selective persistence: a future must remain effective after its cue disappears, but it must lose control when the situation that justified it has changed.
This chapter examines the neural activity that spans such intervals. The psychological literature usually places much of this behavior under working memory. That term remains useful, but it must not be mistaken for the name of a neural organ or a single physiological mechanism.
38.2 Working memory is a psychological construct
The phrase working memory belongs first to psychology. It names a family of behavioral facts. Information that is no longer supplied by the current stimulus can remain available for a short interval; it can support comparison, rule use, calculation, comprehension, or action. People differ reliably in performance on tasks designed to measure that availability, and the construct can organize useful predictions at the psychological level.
None of this requires a one-to-one neural counterpart. Delayed saccades, digit span, mental arithmetic, sentence comprehension, the n-back task, and following a briefly stated rule are all called working-memory tasks. They differ in what must be represented, what transformation must be performed, what response is prepared, how interference is introduced, and whether any item must remain in the focus of attention. A term that groups these behavioral demands does not thereby identify one neural buffer shared by all of them.
The error is subtle because the grammar encourages it. Once working memory becomes a noun, it is easy to ask where the brain stores it, how large its capacity is, or which region serves as its executive. The question has already converted a behavioral description into a neural object. Activation in a working-memory task is then interpreted as activation of working memory itself, and the construct appears to have been localized because it was built into the task label.
The appropriate analysis moves in the other direction. A task should be decomposed into operations and variables that nervous tissue can implement. What information is absent from current input? Which sensory or mnemonic representation must remain consequential? Is the subject monitoring a location, retaining an identity, expecting an event, applying a rule, preparing a movement, estimating elapsed time, or protecting one representation from interference? Which parts of the task require sustained activity, altered synaptic efficacy, recurrent interaction, or a change in the gain applied to later input?
A psychological construct can be valid without being a neural module. The neural explanation begins when the task is decomposed rather than when the construct is assigned to cortex.
This distinction does not make the physiological evidence disappear. It makes that evidence more important. Lateral frontal activity recurs whenever task-relevant information, attention, expectation, or preparation must remain effective over time. The convergence appears in monkey single units, human lesion studies, slow cortical potentials, and functional imaging. The central empirical fact is robust. Its interpretation is broader than a dedicated store called working memory.
At the psychological level, working memory summarizes regularities in behavior: recent information can remain available, capacity is limited, interference matters, and temporary availability predicts performance in many tasks. Those are legitimate phenomena.
At the neural level, the same tasks recruit different combinations of sensory cortex, parietal cortex, hippocampal and medial-temporal systems, frontal cortex, thalamus, basal ganglia, and motor systems. The mapping is many-to-many. A delayed-response task does not isolate a mental substance called working memory, and a region activated by that task does not become its anatomical home.
The useful bridge between levels is an explicit account of the task. Psychological constructs describe organized behavior. Neural mechanisms explain how particular variables remain available, selected, transformed, and connected to action in particular circumstances.
38.3 The delay period made an absent variable visible
The strongest physiological evidence begins before the term working memory became dominant. In 1971, Joaquín Fuster and Garrett Alexander recorded neurons in monkey prefrontal cortex and the mediodorsal thalamus during a delayed-response task. Many neurons changed their firing when the cue appeared, and some remained above their intertrial firing rate throughout the delay [@fusteralexander1971delay]. Activity in frontal and thalamic neurons therefore bridged the interval between information and action. The earliest observation already implicated a recurrent frontothalamic system rather than an isolated cortical container.
Funahashi, Bruce, and Goldman-Rakic later made the represented variable much more precise. In the oculomotor delayed-response task, a monkey fixated a central point while a peripheral target appeared briefly. The target disappeared, fixation continued through a blank delay, and the disappearance of the fixation point instructed a saccade to the remembered location. Retinal stimulation and eye position were controlled during the delay, and the direction and timing of the eventual response could be measured exactly [@funahashietal1989mnemonic].
Neurons around the principal sulcus showed spatially selective delay activity. One neuron might remain active after a cue in the upper-left visual field but not after an otherwise identical cue in the lower-right field. The preferred region was called a memory field. Across the population, the pattern of activity contained information about a location that the retina no longer supplied.
The activity was not merely correlated with successful performance. Lesions in and around the principal sulcus produced delay-dependent errors for remembered locations in the contralateral visual field while leaving visually guided saccades substantially intact. Errors increased as the blank interval lengthened [@funahashietal1993lesions]. The relevant frontal tissue was therefore causally necessary for normal performance when the cue had to remain effective across time.
These experiments established a firm result: prefrontal population activity can preserve task-specific information after the eliciting stimulus has disappeared, and disruption of the relevant tissue impairs the behavior that depends on that information. They did not establish that working memory, as a psychological faculty, had been found inside the principal sulcus.
The task itself contains more than retrospective memory. The cued location is also the location to which attention remains directed and the target of a future saccade. Later antisaccade experiments separated the remembered cue location from the required movement direction. Most delay activity followed the location of the cue, while a smaller population followed the direction of the forthcoming saccade [@funahashietal1993antisaccade]. The physiology contained both retrospective and prospective variables rather than respecting the boundary of one psychological construct.
Persistent firing should therefore be described without either minimizing or personifying it. It is not a neuron “holding an image in mind.” It is sustained, selective activity through which an absent variable continues to alter the state of a circuit and the probability of a later action.
38.4 The frontal regularity is broader than memory
A second experimental tradition reached the same general problem without asking subjects to retain a visual item. Walter and colleagues presented a warning stimulus followed, after a predictable interval, by an imperative stimulus that required a response. A slow negative voltage developed over the scalp during the interval. They called it the contingent negative variation, or CNV, and described it as an electrical sign of association and expectancy [@walteretal1964cnv].
The CNV occurs while the organism is waiting in a task-defined state. The warning signal has changed what the next event means, when it is expected, and which response may be required. Anticipatory attention, temporal expectancy, motivation, and motor preparation all contribute. There need be no vanished object stored in a buffer. What persists is a relation among the warning cue, the expected event, the task rule, and the prepared response.
Lesion and intracranial evidence place frontal cortex within this sustained preparation without assigning the potential to one source. Patients with prefrontal lesions showed a marked reduction of the late CNV [@rosahlknight1995cnv]. Subdural recordings revealed CNV activity at multiple cortical sites, including prefrontal, supplementary sensorimotor, primary sensorimotor, temporal, and occipital regions [@hamanoetal1997cnv]. The scalp waveform is a distributed field phenomenon. Its frontal contribution reflects the maintenance of an anticipatory task configuration, not a memory trace measured from outside the skull.
Functional imaging supplied a third scale of observation. In a delayed face task, Courtney and colleagues separated transient responses to visual events from activity sustained across the blank interval. Sustained activity appeared in a distributed system that included frontal and posterior visual regions [@courtneyetal1997sustained]. The result agrees with the single-unit and slow-potential evidence: when behavior depends on a relation that must persist across time, frontal activity often persists with it.
The convergence across methods should not be explained away. Single neurons, field potentials, lesions, and BOLD measurements all implicate frontal systems in temporally extended task organization. Nor should the convergence be translated automatically into “the neural basis of working memory.” The same frontal regularity appears during sustained attention, expectancy, monitoring, response preparation, and rule maintenance.
Lebedev and colleagues made the distinction explicit by requiring monkeys to remember one location while attending to another. Although the task imposed a substantial short-term memory demand, the largest proportion of spatially tuned prefrontal neurons represented the attended location rather than the remembered one [@lebedevetal2004attention]. Delay activity could not be understood as storage alone. It reflected which variable currently had priority within the task.
The neural regularity is broader than the psychological category. Frontal systems remain engaged when a task relation must continue to organize behavior across an interval, whether the relation is described psychologically as memory, attention, expectancy, or preparation.
38.5 What remains effective during a delay
The word maintenance can also conceal important differences. During one delay, the system may preserve the location of an absent target. During another, it may keep a conditional rule effective, sustain attention to one sensory stream, prepare one of several possible actions, estimate when an event will occur, or retain an intermediate result that will be transformed at the next step. These variables can coexist, and a single neuron can respond to combinations of them.
A useful mechanistic term is control state. The term does not name another faculty. It describes a measurable configuration of the system that changes how later input will be processed and which action it will favor. A red light can mean stop, press left, compare the next two images, or ignore the next event. The visible stimulus is the same. The maintained rule changes the mapping from input to action.
Control states need not be represented by an unchanging set of neurons firing at one rate. In monkeys applying different rules to the same visual choices, an instruction cue drove prefrontal populations through a sequence of states before activity settled into a rule-specific configuration. When the choice stimulus arrived, the population response evolved differently according to the maintained rule [@stokesetal2013dynamic]. The rule was expressed in the geometry and trajectory of population activity, not in a dedicated “rule cell” that remained uniformly active.
The same distinction applies to retained content. Individual prefrontal neurons can change their tuning across a delay while the population preserves a stable representation in a lower-dimensional subspace [@murrayetal2017stable]. Stability at the behavioral level therefore does not require every cellular component to remain static. A moving population can preserve an invariant that downstream systems can read.
Nor must every potentially relevant item remain continuously decodable at the same strength. In human experiments, an item removed from the current focus of attention could become difficult to decode from ongoing activity and later return to an active pattern. A pulse of transcranial magnetic stimulation delivered during that latent interval temporarily reinstated a stimulus-specific pattern and selectively altered later performance [@roseetal2016latent]. The result shows that temporary availability can depend on changed network state as well as continuous high firing.
These findings do not replace one working-memory mechanism with another. They dissolve the demand for one mechanism. Sustained spiking, dynamic population trajectories, altered effective connectivity, short-lived synaptic changes, and later reactivation can all allow an absent variable to influence behavior. Which mechanism dominates depends on the content, delay, interference, expected use, and species.
The distinction between psychological and neural description can now be stated precisely:
Working memory is the behavioral label for temporary availability. The neural account is an inventory of the particular states, representations, transformations, and recurrent interactions that make information effective in a particular task.
38.6 Holding the line is a property of loops
The earliest delay recordings included the mediodorsal thalamus, but the later history often made persistent firing sound like a local achievement of prefrontal microcircuits. Causal circuit experiments have restored the loop.
In a mouse delayed nonmatching task, pathway-specific perturbation showed that projections from mediodorsal thalamus to medial prefrontal cortex were required during the maintenance interval, whereas the return projection from prefrontal cortex to mediodorsal thalamus contributed especially to the subsequent choice. Delay-active prefrontal neurons lost their sustained firing when mediodorsal input was suppressed, and enhancing mediodorsal excitability improved performance [@bolkanetal2017thalamic].
A related result emerged in a delayed motor task. Neurons in anterior lateral motor cortex and connected thalamic nuclei both carried preparatory activity. Silencing either node caused activity in the other to collapse and impaired the future movement [@guoetal2017loop]. Persistent activity was sustained by reciprocal excitation across regions. The interval-spanning state belonged to the loop.
Basal-ganglia circuits contribute another requirement: not every available representation should enter effective control. In a human imaging experiment, activity in prefrontal cortex and basal ganglia before a memory array predicted how effectively irrelevant items would be excluded from later representation [@mcnabklingberg2008access]. The result is often described as gating access to working memory. Mechanistically, it concerns which information is admitted to a protected, behaviorally effective state.
Posterior sensory and parietal systems carry properties of the absent stimulus and the spatial or attentional relations needed for the task. Hippocampal and medial-temporal systems contribute relations, sequences, and context when the delay depends on more than one item or event. Premotor and motor systems carry prospective actions. Neuromodulatory systems alter the stability and gain of the participating populations. No frontal signal can guide behavior unless it changes processing in these connected systems.
The regulated body also remains inside the loop. The selected water source in the opening example does not retain a fixed value while a neutral memory system carries it forward. Hypothalamic, insular, brainstem, orbital, ventromedial, and striatal signals continue to report and predict bodily state, cost, risk, and outcome. A goal can weaken because the deficit has been corrected, strengthen because the predicted heat load has increased, or be abandoned because another threat has become more consequential. Maintenance protects an expected outcome from irrelevant competition; it does not quarantine the outcome from revaluation.
The comparative evidence reinforces the functional point. Pigeons can be instructed on each trial either to remember a stimulus or to forget it. Neurons in the nidopallium caudolaterale sustained activity through the delay after a remember cue and reduced that activity after a forget cue; later performance followed the same instruction [@rosecolombo2005avian]. The avian pallium and mammalian prefrontal cortex are organized differently. Both participate in circuits that regulate whether absent information remains behaviorally available.
The control problem is therefore older than primate granular cortex. Evolution has repeatedly built systems in which current activity is shaped by a cue whose consequence lies in the future.
38.7 Stable enough to persist, flexible enough to change
The title Holding the Line can now be understood without assigning a line-holder to the brain. A task cue changes the state of a distributed controller. That state increases the influence of some representations and actions, decreases the influence of others, and persists long enough for the relevant consequence to be reached.
The central engineering problem is not maximal persistence. It is deciding which new events count as disturbances and which count as evidence that the state itself is wrong. A passing sound may be irrelevant to the route toward water. A predator on the route is not. The first should be rejected; the second should reorganize behavior.
Several operations traditionally grouped under executive function contribute to this balance, but they should not be collapsed into one faculty.
Maintenance keeps a rule, target, or expected outcome consequential after its cue disappears. Proactive interference control configures the system in advance so that an irrelevant stimulus or prepotent response has less influence. These operations are closely related because a strongly represented rule biases competition before conflict occurs.
Reactive stopping solves a different problem. An action that is already moving toward execution must be interrupted when a stop signal arrives. Damage to right inferior frontal cortex impairs stop-signal performance. Functional imaging and tractography place this region in a stopping network that also includes presupplementary motor cortex and the subthalamic nucleus [@aronetal2003stopping; @aronetal2007network]. Stopping is not merely the negative face of remembering a rule.
Updating also differs from maintenance. New evidence must be identified as relevant, the current state must lose priority, and a replacement must gain access to control. A system that updates too readily becomes distractible. A system that updates too slowly perseverates. The two errors lie on opposite sides of the same stability–flexibility problem.
Planning requires several maintained states to be related hierarchically. An overarching outcome must survive while intermediate subgoals are installed and released. This is more than keeping one item active for a longer delay. It requires monitoring progress, representing transitions, and revising the route when an intermediate prediction fails.
Initiation should remain separate again. A goal can be represented, valued, and maintained without mobilizing the organism to begin. The final chapter of this unit will examine apathy, abulia, and akinetic mutism precisely because making a course of action available is not the same as initiating it.
The result is not a new list of modules. It is a decomposition of control problems. The same frontal, thalamic, striatal, posterior, and motor systems can participate in several operations, but participation does not make the operations identical.
Maintained constraints usually reduce an enormous response space to actions that fit the task. In an unusual problem, that advantage can become a cost.
Reverberi and colleagues asked patients with focal lesions and healthy comparison participants to solve matchstick equations. The most difficult problems required changing the ordinary interpretation of the symbols. Patients with lateral frontal lesions solved 82% of these problems, compared with 43% of the comparison group [@reverberietal2005insight].
The result does not show that lateral frontal damage improves reasoning in general. It demonstrates a bounded tradeoff. A system that strongly constrains interpretation toward normally useful representations can impede a solution that requires abandoning those representations. Stability is valuable because the world is usually structured. Flexibility is valuable because the current structure is sometimes wrong.
38.8 When a known rule loses control
Goal neglect provides the clearest separation between possessing information and allowing that information to govern behavior. A person can understand an instruction, repeat it accurately, and then fail to implement one of its requirements when several demands must be coordinated. The rule remains available for report but drops out of effective control [@duncanetal1996goal].
This dissociation directly challenges the idea that a remembered rule is simply stored or absent. The relevant question is what the representation can do. A sentence repeated to the examiner has psychological availability. It has behavioral control only when it changes how later events are interpreted and which actions are selected.
Novel and open-ended situations make the distinction especially visible. Shallice and Burgess described patients who performed many conventional tests adequately but failed when they had to organize several simple activities over an extended interval with little external structure [@shalliceburgess1991strategy]. The component skills were present. The patients did not reliably construct and sustain a task model that coordinated them.
Goal neglect is therefore not ordinary forgetting, and it is not evidence that one executive has disappeared. It is a failure of a known requirement to achieve or retain the neural influence needed to organize the rest of the task.
38.9 When the present gains too much control
Distraction reveals the same problem from the other side. During a delayed-response task, an irrelevant event can reorganize sensory, affective, and frontal activity strongly enough to degrade the state that supports the pending response.
Dolcos and McCarthy presented emotional or neutral distractors during the delay of a face working-memory task. Emotional distractors impaired later performance. They increased activity in the amygdala and ventrolateral prefrontal cortex while producing relative reductions in the dorsolateral prefrontal and lateral parietal activity sustained by the task [@dolcosmccarthy2006interference]. The effect was a redistribution across interacting systems, not a contest in which a ventral valuation module pulled a dorsolateral memory module offline.
The distinction between value and salience developed in Chapter 37 matters here. A disturbing image can command processing because it is biologically or emotionally significant without offering a positively valued outcome. Emotional distraction is not a competing bid in a reward currency. It is an event that changes processing priority and can interfere with the configuration required by the ongoing task.
The clinical extreme is utilization behavior. Lhermitte described patients who grasped and used objects placed before them although the objects were irrelevant to any stated goal. A comb invited combing; spectacles were put on; a glass invited drinking [@lhermitte1983utilization]. The behavior exposed an ordinary fact in exaggerated form: visible objects continuously offer actions. In intact control, those affordances compete with rules, social context, predicted outcomes, and longer-range goals. After frontal-network injury, the immediate invitation can acquire disproportionate control.
Utilization behavior should not be labeled a pure dorsolateral deficit or a simple release from one inhibitory center. The syndrome has followed different frontal and frontal-subcortical injuries, including paramedian thalamic infarction [@lhermitte1983utilization; @eslingeretal1991thalamic]. It is best understood as environmental dependency: the present scene governs action more strongly than the absent variables that would normally place the scene in context.
38.10 When an old state will not release
Perseveration produces the opposite clinical appearance. The patient continues to sort, search, speak, or act according to a rule that no longer fits the evidence. The problem is not failure to maintain a control state. It is failure to replace one.
Brenda Milner’s card-sorting studies established a strong relationship between frontal lesions and difficulty shifting from one sorting principle to another [@milner1963cardsorting]. The Wisconsin Card Sorting Test that grew from this tradition remains clinically useful because it exposes the interaction of rule discovery, feedback use, maintenance, and switching. It is not an anatomical scanner for dorsolateral prefrontal cortex. Poor performance can arise from several component failures and from damage outside one frontal territory.
The contrast with goal neglect is instructive. In goal neglect, a valid rule loses effective control. In perseveration, an invalid rule retains it. Distractibility and environmental dependency make behavior too sensitive to the present. Perseveration makes behavior insufficiently sensitive to change. No single score called executive function captures these opposed failures.
Stability and flexibility are not competing explanations of frontal function. They are the opposing demands that any temporally extended controller must regulate.
38.11 Coda: the future enters the present body
An absent future can govern action because the nervous system does not return to its pre-cue state when the cue disappears. Frontal, thalamic, striatal, posterior cortical, hippocampal, motor, and bodily systems remain configured by what the cue established. Their ongoing activity and altered interactions keep selected information, rules, priorities, and prospective actions effective across time.
Psychology describes many of these behaviors with the construct working memory. The construct is useful because temporary availability is a real feature of behavior. The physiology does not reveal a working-memory organ. It reveals multiple ways in which a distributed controller can preserve and transform variables that are absent from current input.
The control state must remain open to the body. A predicted outcome can change heart rate, vascular state, visceral sensation, and readiness for action before the outcome arrives. Those anticipatory changes can in turn alter what the outcome is worth and how urgently it should be pursued. Maintenance explains how a future remains behaviorally effective. It does not yet explain how that future becomes embodied in the present.
The next chapter, Feeling the Future, follows that prediction into autonomic and interoceptive state.
Established findings. Neurons in primate lateral prefrontal cortex can show selective activity across delays, and lesions of the relevant tissue produce delay-dependent behavioral deficits. Sustained frontal involvement also appears in the human CNV and in functional imaging when attention, expectancy, task rules, or stimulus information must remain effective across time. Delay-period activity can represent remembered information, attended information, rules, task context, and prospective actions. Persistent states depend on recurrent interactions with thalamus and other cortical and subcortical systems. Goal neglect, distraction, reactive stopping, perseveration, and environmental dependency are dissociable control failures. Related remember-versus-forget regulation occurs in the differently organized avian pallium.
Working synthesis. Working memory is a useful psychological construct, not the name of a neural module. The recurring physiological problem is maintained control across time. Frontal systems contribute population states that preserve, prioritize, and transform task-relevant variables; thalamic and striatal loops help stabilize, gate, and update those states; posterior, hippocampal, sensory, and motor systems carry much of the content and action structure; bodily systems continue to revalue the expected consequence. The resulting controller is stable when disturbance is irrelevant and revisable when evidence changes the state.
Open questions. Different tasks use different mixtures of sustained spiking, dynamic population coding, altered synaptic or connectivity states, and later reactivation. The conditions that favor each mechanism remain under study. It is also unresolved how cortical, thalamic, and striatal loops classify a new event as interference to be rejected or evidence that requires updating; how several nested goals remain coordinated; and how maintenance changes when the relevant variable is spatial, verbal, social, bodily, or extended across minutes rather than seconds. These questions concern the mechanisms of temporary availability. They do not require a single neural entity corresponding to the psychological name working memory.