Jared Edward Reser, PhD
Abstract
Theories of working memory have traditionally focused on how information is maintained, prioritized, removed, retrieved, or protected from interference. A complementary question concerns what retained information does during the generation of successive cognitive states. The iterative updating model proposes that thought proceeds through a recurrent sequence in which some working-memory representations are retained, others are removed, and newly generated representations are added. Retained representations are not treated as passive residues of previous processing. Rather, they contribute to the selection and interpretation of subsequent contents. Newly recruited contents then join the retained configuration, alter the effective cognitive context, and participate in generating later states. This recurrent retain-add-reweight process has been described in terms of state-spanning coactivity (SSC), incremental change in state-spanning coactivity (icSSC), multiassociative search, polyassociativity, iterative compounding, and progressive imagery modification (Reser, 2016, 2022-2026). The present article evaluates the empirical status of this architecture.
Converging behavioral, neurophysiological, neuroimaging, intracranial, and perturbational findings support many of its required operations. Working memory can undergo item-specific removal while preserving other contents; mnemonic resources can be reallocated from obsolete to newly relevant information; prefrontal representations can remain active for relevance-dependent intervals; retained representations can change their coding and behavioral priority when new information arrives; multiple preceding cues can jointly influence subsequent processing; retrieval history and reinstated neural context can predict endogenous recall; latent representations can remain recoverable after leaving the focus of attention; mental imagery can be progressively transformed and can generate information not explicitly supplied in advance; and maintained neural activity can causally bridge temporally separated events. The strongest evidence therefore supports selective partial updating and context-dependent successor generation as real cognitive operations. The principal unresolved question is whether several independently identifiable, selectively retained representations can be shown to jointly generate an endogenous successor that subsequently joins them and contributes to generating another successor. Demonstrating this repeated causal sequence would provide direct architectural evidence that iterative updating is not merely a property of working memory, but a generative organization of thought.
Keywords: working memory, thought, iterative updating, state-spanning coactivity, multiassociative search, associative retrieval, mental imagery, cognitive architecture, neural dynamics, reasoning
1. Introduction
Working memory is commonly described as a limited-capacity system that temporarily maintains information in a state suitable for ongoing cognition. This characterization has generated extensive research on capacity, maintenance, interference, attention, gating, updating, removal, retrieval, and neural persistence. Yet these literatures leave open a more general architectural question: How does the information that remains available from one cognitive moment participate in producing the next cognitive moment?
The iterative updating model proposes an answer. Rather than treating each cognitive state as an isolated configuration that is followed by another, the model describes thought as a recurrent transformation of a partially preserved representational state. At one moment, working memory may contain representations B,C,D,E. Following an update, B may be removed while C,D,E remain and a new representation F becomes active. A further transition may remove D, preserve C,E,F, and add G:
\{B,C,D,E\}
\rightarrow
\{C,D,E,F\}
\rightarrow
\{C,E,F,G\}.
The essential claim is not simply that consecutive states overlap. The retained representations are proposed to participate in selecting the addition. Once recruited, the addition becomes part of the configuration responsible for the next search. The cognitive state is therefore simultaneously a product of preceding processing and a set of conditions governing subsequent processing. This operation is central to features 4 through 7 of the model’s eight-feature summary: item activity is staggered and overlapping; active contents provide search parameters; newly activated contents join those that remain; and the resulting search is a modified iteration of the preceding search rather than an independent restart.
Thought iterative updating(20260907-214548).pdf
The earlier formulation described representations persisting between states as exhibiting state-spanning coactivity (SSC), and the changing membership of the coactive configuration as incremental change in state-spanning coactivity (icSSC). It further proposed polyassociativity, subsequently termed multiassociative search, whereby simultaneously available representations pool their influence in recruiting a subsequent representation. The recruited representation then enters the same state-space from which subsequent associations are generated (Reser, 2016). The 2016 formulation explicitly states that new representations join those that recruited them and are incorporated into later searches.
Incremental change in state spanning cortical.pdf
The present article asks how much of this architecture is already supported empirically. The answer is more substantial than might initially appear. Research literatures that developed largely independently have established selective updating, relevance-dependent maintenance, latent working-memory states, contextual retrieval, multiple-cue facilitation, representational transformation, mental-image manipulation, temporal bridging, and working-memory-dependent learning. The important task is to determine which findings merely establish required components, which demonstrate interactions among those components, and which approach the stronger claim that retained representations repeatedly generate successors during an unfolding train of thought.
The empirical literature does not yet contain a single experiment demonstrating the complete sequence during spontaneous cognition. However, almost every required operation has meaningful independent support, and several experiments connect two or more proposed mechanisms within the same task. The crucial remaining gap is the repeated chain:
\text{retained contents}
\rightarrow
\text{endogenous successor}
\rightarrow
\text{updated retained set}
\rightarrow
\text{next endogenous successor}.
The evidence review underlying the present synthesis reached precisely this conclusion.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
The argument developed here is therefore neither that iterative updating has already been conclusively demonstrated as the universal mechanism of thought nor that its components are merely speculative. The stronger position is that existing evidence supports a convergent architecture whose distinctive coordination can now be stated precisely enough to test.
2. Iterative Updating as a Generative State-Transition Architecture
The concept of updating is often used broadly. Information can be updated by adding an item, overwriting a memory, replacing an old representation with a new one, retrieving something previously inactive, or revising the priority assigned to already represented information. Iterative updating specifies a narrower organization.
Let the currently operative representational state be:
S_t=\{r_1,r_2,\ldots,r_k\}.
During an update, some subset D_t becomes obsolete or insufficiently relevant and is removed. Some subset R_t remains available, where:
R_t=S_t\setminus D_t.
A newly generated representation x_{t+1} is selected conditional on the operative state:
x_{t+1}\sim P(x\mid S_t,G_t,L_t),
where G_t denotes current goals or task demands and L_t represents information embedded in long-term memory and learned network structure. The next working state becomes:
S_{t+1}=R_t\cup\{x_{t+1}\}.
Critically,
x_{t+2}\sim P(x\mid S_{t+1},G_{t+1},L_{t+1}).
The product of one search therefore becomes a parameter of the next search.
This is iterative compounding. Cognitive products can accumulate because an intermediate result does not have to terminate the computation that generated it. It can instead become part of the conditions under which the next intermediate result is produced.
The model’s multiassociative-search algorithm adds another feature. A newly entering representation can redistribute influence within the retained state. Some retained items become more relevant in light of the addition, others become less relevant, and the effective representation of an item may change according to its context. The model therefore does not require working-memory items to remain immutable objects occupying fixed slots. Thought iterative updating(20260907-214548).pdf Figure 28 makes this explicit by proposing that a new item redistributes activation across previously available contents, changing their relative contribution to successor selection.
Thought iterative updating(20260907-214548).pdf
This yields a more complete transition:
(S_t,W_t)
\rightarrow
x_{t+1}
\rightarrow
(S_{t+1},W_{t+1}),
where W_t represents the momentary weighting, priority, or effective influence of individual contents.
The theoretical unit of interest is consequently not the isolated item and not merely the transition between two items. It is the evolving configuration.
3. Selective Partial Updating Is Empirically Well Supported
One of the least controversial requirements of iterative updating is that information can be selectively removed while other information remains available. Behavioral research now supports this operation directly.
Ecker, Oberauer, and Lewandowsky (2014) tested whether working-memory updating requires wholesale replacement or whether individual obsolete items can be removed selectively. Their experiments supported item-specific removal. Participants could discard information that was no longer required while preserving other contents.
This is structurally similar to:
\{B,C,D,E\}
\rightarrow
\{C,D,E,F\},
because the previous state does not disappear in its entirety. Selected components survive while another component is replaced.
Taylor, Tomić, Aagten-Murphy, and Bays (2023) approached the issue from a resource perspective. Participants maintained visual information and were instructed under different conditions to retain, repeat, add, or replace items. Their results were consistent with mnemonic resources being withdrawn from obsolete information and reallocated to replacement items.
Together, these studies establish two operations required by iterative updating: selective removal and reallocation to newly relevant content.
A particularly important neurophysiological result was reported by Sawagashira and Tanaka (2025). Three macaques performed a rule-dependent oculomotor n-back task in which spatial memories remained relevant for different durations depending on the current rule. Recordings from 152 lateral prefrontal neurons identified populations carrying directional memory information as well as neurons showing transient signals when particular memories ceased to be required. Sixty neurons exhibited directional memory signals and 61 exhibited directional extinction-related signals. Memory representations therefore did not simply decay according to how long ago a cue appeared. Their duration was sensitive to whether the represented information remained behaviorally necessary.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
This result is especially relevant to the model’s contention that replacement need not obey a rigid first-in-first-out schedule. An older item can persist while a newer item leaves if the older item remains useful.
Sawagashira and Tanaka also provided causal evidence. Neural population activity predicted future choices, and electrical stimulation at a subset of recorded sites produced condition-specific errors when particular memories needed to be retained.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
These findings do not demonstrate spontaneous multiassociative thought because the information and updating requirements were externally imposed. They do, however, show that the nervous system can perform relevance-sensitive partial updating using content-bearing neural states.
The retain-remove component of iterative updating should therefore be regarded as strongly supported at the functional level.
4. Representational Persistence Does Not Require an Immutable Neural Code
The original SSC formulation emphasized sustained neural firing. Such firing remains an important part of working-memory physiology, but current evidence suggests that representational persistence should be distinguished from literal persistence of exactly the same cellular activity pattern.
Kamiński and colleagues (2017) recorded individual neurons in humans performing working-memory tasks. Persistent neural activity was observed in medial frontal and medial temporal regions. In the hippocampus and amygdala, persistent responses could carry information about particular remembered stimuli, demonstrating that content-specific information can remain available in sustained neuronal activity.
Yet other evidence indicates that stable information can be represented through changing population dynamics. Murray and colleagues (2017) analyzed macaque prefrontal activity and found that stable population-level coding of working-memory information could coexist with substantial temporal heterogeneity at the level of individual neurons. The representation could remain stable even though the particular cellular pattern implementing it evolved.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
This distinction is important because three propositions should not be conflated:
\text{information remains available}
\Downarrow
\text{a population representation remains recoverable}
\not\Rightarrow
\text{the identical neurons fire continuously}.
Iterative updating primarily requires the first proposition. A strict cellular interpretation of SSC requires the third.
The theoretical architecture can therefore be strengthened by treating continuous sustained firing as one implementation of state-spanning representation rather than its universal definition.
4.1 Latent state-spanning availability
Research on unattended working-memory contents reinforces this conclusion.
Lewis-Peacock, Drysdale, Oberauer, and Postle (2012) showed that the neural signature of an item could decline substantially after it left the focus of attention, even though the information remained relevant and behaviorally recoverable.
Rose and colleagues (2016) subsequently demonstrated that transcranial magnetic stimulation could transiently restore a decodable representation of information that had become difficult to identify from ongoing activity but remained relevant to the task. The perturbation also affected behavior.
Wolff and colleagues (2017) used EEG and an impulse-response method to reveal information stored in hidden working-memory states. Information could therefore remain functionally present even when it was not expressed as straightforward persistent activity.
Together these findings strongly support a distinction between actively expressed and latent but recoverable states.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
A useful extension of the terminology is consequently:
Active state-spanning representation: information persists through ongoing population activity.
Latent state-spanning availability: information remains encoded in a temporarily altered neural state and is rapidly recoverable, even when its content is not continuously expressed in overt firing.
The architecture requires that retained information remain capable of influencing later computation. It need not require all retained contents to occupy the same neurophysiological state.
5. New Information Changes the State It Enters
Partial updating would be cognitively limited if retained representations simply remained unchanged while new representations accumulated beside them. The iterative updating model instead proposes reciprocal influence. Newly recruited information changes the effective weighting and possibly the neural composition of retained representations.
Warden and Miller (2007) provide unusually direct evidence for this proposition. They recorded lateral prefrontal neurons while macaques remembered sequences of two objects. Neural representations did not behave as though two independently encoded objects were simply stored beside one another. The identity of the second object influenced how the first was represented.
This finding corresponds closely to the proposed transition from a set such as:
\{B,C,D,E\}
to:
\{C,D,E,F\},
where F not only enters the set but modifies the effective cognitive meaning of the configuration.
The empirical review found the correspondence to the model’s Figure 28 particularly strong. The figure proposes that a newly entering item redistributes influence among already active items and changes their contribution to the following search. Warden and Miller demonstrate the physiological plausibility of the first part of this process, although they did not demonstrate that the resulting state subsequently generated a third internally produced representation.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
Panichello and Buschman (2021) provide complementary population-level evidence. During working-memory and attentional-selection tasks, frontoparietal representations transformed according to behavioral priority. Information moved into representational formats better suited for controlling subsequent behavior.
These results motivate an important principle:
Retention does not imply representational immutability.
A concept can remain functionally continuous across states while its neural expression, priority, relational meaning, or downstream influence changes according to the other information with which it is coactive.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
This point strengthens the original model’s description of neural ensembles as fuzzy and context-sensitive rather than fixed.
6. Multiassociative Search: Can Several Representations Jointly Constrain a Successor?
The more distinctive claim of iterative updating concerns successor selection.
A simple associative chain can be written:
A\rightarrow B\rightarrow C\rightarrow D.
Multiassociative search instead proposes transitions more like:
\{A,B,C,D\}\rightarrow E.
Several simultaneously available representations constrain which candidate becomes most strongly activated.
The longer formulation describes this as a cooperative search through long-term memory in which coactive representations spread their combined influence and converge on an associated addition. The new item then becomes part of the next search configuration.
Thought iterative updating(20260907-214548).pdf
Evidence relevant to this process comes from several research traditions.
6.1 Multiple-prime summation
Balota and Paul (1996) examined whether multiple semantic primes could contribute jointly to target processing. Across several experiments, two relevant primes could facilitate a related target more strongly than one alone. In many conditions, the contributions were approximately additive.
This establishes an important prerequisite: associative activation need not be determined by one preceding representation.
However, a presented target is not equivalent to a self-generated thought. The experiment demonstrated that multiple inputs can facilitate target recognition, not that a retained configuration autonomously selected the target.
6.2 Configuration-sensitive integration
Lavigne and colleagues (2016) found that priming could depend on a learned higher-order configuration rather than being completely reducible to separate pairwise relationships between primes and targets.
This result is relevant because multiassociativity predicts that a combination of representations can carry information unavailable from any member considered alone.
The distinction between learned configurations and novel convergence events remains important. The 2016 model specifically proposed that familiar individual associations might converge successfully even when the complete conjunction of cues had never previously occurred.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
That prediction deserves direct experimental testing.
7. Compound Cuing Brings the Evidence Closer to Endogenous Thought
Lohnas and Kahana’s (2014) work on compound cuing in free recall is especially important because the successor representation is generated internally rather than supplied by the experimenter.
Their analyses asked whether previous retrieval history influences the next recall beyond the most recently recalled item. Evidence indicated that more than the immediately preceding retrieval could affect what was recalled next and how quickly it was retrieved.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
This begins to approximate:
\{C,D\}\rightarrow E
rather than:
D\rightarrow E.
That is a meaningful advance over ordinary semantic priming because the transition occurs within a self-generated memory sequence.
However, compound cuing does not uniquely establish multiassociative updating. Retrieved-context theories predict related phenomena. Each retrieved item can reinstate temporal context, which in turn changes the context that cues subsequent retrieval. Thus, a distributed contextual signal can preserve information about several previous events without requiring several individually maintained semantic representations.
This alternative is theoretically important because it identifies the level at which iterative updating must become more precise.
The distinctive prediction is not merely:
P(E\mid C,D)>P(E\mid D).
It is that independently measurable contents that remain selectively relevant should explain successor selection beyond recency, fixed task state, and retrieved temporal context.
8. A Boundary Condition: Semantic Compound Cuing Is Not Universal
A strong theory should incorporate findings that constrain its scope.
Morton and Polyn (2016) investigated semantic organization during free recall and found relatively little support for the proposition that several previous semantic recalls invariably accumulate into a compound semantic cue. Although recall was semantically organized, transitions were often best predicted by the semantic relationship to the immediately preceding item.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
This result argues against the strongest possible interpretation of polyassociativity:
Every recently activated representation materially contributes to every subsequent association.
That formulation is unnecessary.
The iterative updating model already includes selective retention, relevance-dependent removal, and unequal influence among coactive contents. A more precise prediction is therefore:
Multiple representations jointly constrain successor selection when they remain functionally prioritized components of the operative cognitive state.
The distinction is substantial.
A memory that was encountered recently but has become irrelevant should exert relatively little influence. A representation encountered earlier that remains essential to the current problem may continue exerting strong influence.
This leads to one of the clearest discriminating predictions generated by the present synthesis:
\text{relevance} > \text{recency}
under conditions requiring continued constraint maintenance.
The model therefore predicts stronger multiassociative effects in structured reasoning, planning, imagery, and problem-solving tasks than in unconstrained free recall.
9. Neural Context Predicts the Direction of Endogenous Retrieval
Intracranial recordings provide another bridge between retained internal state and self-generated succession.
Manning and colleagues (2011) examined neural context reinstatement during memory search. Activity preceding a recall reinstated aspects of temporal context associated with encoding, and the reinstated state predicted temporal organization in subsequent recall.
Manning and colleagues (2012) extended this approach to semantic organization. In 46 neurosurgical participants, spontaneously reactivated neural patterns in frontal and temporal regions preceded recall and predicted semantic clustering.
The findings establish three propositions highly relevant to iterative updating:
- An internal neural context exists during memory search.
- Aspects of that context are reinstated before self-generated retrieval.
- The structure of the internal state predicts where retrieval proceeds next.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
What these studies do not establish is that several independently identifiable working-memory representations jointly produced the subsequent item. A distributed contextual representation can explain much of the same evidence.
The appropriate conclusion is therefore that internal neural state influences endogenous successor selection, while the representational composition of that state remains an open question.
10. Progressive Imagery Modification
Iterative updating was originally proposed not only as a mechanism of semantic retrieval but also as a means of progressively transforming mental imagery.
The model proposes a reciprocal sequence:
\text{retained conceptual constraints}
\rightarrow
\text{imagery construction}
\rightarrow
\text{new information generated within imagery}
\rightarrow
\text{conceptual update}
\rightarrow
\text{revised imagery}.
The longer manuscript describes this as progressive imagery modification. Figure 49 depicts conceptual contents in association networks constraining a sensory image, features of that image activating a new higher-order item, and the revised conceptual set subsequently generating another image.
Thought iterative updating(20260907-214548).pdf
Several empirical literatures support portions of this cycle.
10.1 Internally generated representations can be transformed
Schlegel and colleagues (2013) examined mental construction, maintenance, and disassembly of unfamiliar visual forms. Distributed neural patterns differentiated the starting representation, the manipulation, and the resulting configuration.
Christophel, Cichy, Hebart, and Haynes (2015) showed that working-memory contents remained represented during mental transformations. Participants mentally rotated complex visual patterns, and information concerning both remembered and transformed contents could be decoded from parietal and early visual cortex.
These studies establish that imagery can act as an evolving computational representation rather than merely as a static memory trace.
10.2 Imagery involves strong top-down processing
Dijkstra and colleagues (2017) compared connectivity during visual perception and imagery. Their modeling indicated stronger top-down influences during imagery and stronger bottom-up influences during perception. Imagery vividness was associated with top-down connectivity toward early visual areas.
This finding is consistent with the proposed conceptual-to-sensory direction of progressive imagery modification.
10.3 Imagery can generate information
A particularly relevant behavioral finding comes from Finke, Pinker, and Farah (1989). Participants mentally constructed and transformed visual patterns and could subsequently discover new interpretations of the resulting mental image.
This is important because it demonstrates that imagery can yield information that was not explicitly supplied as part of the original symbolic instruction.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
The result supports a central functional premise of progressive imagery modification: internal sensory construction can contribute novel informational content back to cognition.
The critical missing step is recursion.
Existing studies generally demonstrate a transformation and perhaps a subsequent discovery. The stronger prediction is:
\text{newly discovered feature}
\rightarrow
\text{updated conceptual state}
\rightarrow
\text{next transformed image}.
A multistep task directly measuring this return loop would provide a strong test of the theory.
11. Temporal Bridging, Learning, and the Accumulation of Cognitive Products
A generative architecture of thought must allow representations separated in external time to become jointly relevant internally.
Gilmartin, Miyawaki, Helmstetter, and Diba (2013) provide strong causal evidence for this general principle. During trace fear conditioning, optogenetic silencing of prelimbic prefrontal activity during the interval separating a cue and an outcome impaired formation of the association.
The experiment does not establish a multi-item cognitive workspace. It does show that activity spanning the temporal gap is functionally necessary for linking nonoverlapping events under the tested conditions.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
This corresponds closely to the SSC proposal that sustained internal representations can make two externally nonconcurrent events simultaneously available to neural learning mechanisms.
Working-memory processing also influences long-term memory formation. Ranganath, Cohen, and Brozinsky (2005) showed relationships between neural activity during working-memory maintenance and subsequent long-term memory.
Sabo and Schneider (2025) similarly found that processing information in working memory improved later long-term-memory retrieval. Their experiments provide modern evidence that what occurs within temporary processing states can modify later memory.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
These findings support the broad transition:
\text{co-processing}
\rightarrow
\text{changed future accessibility}.
They do not establish the more specific synaptic hypothesis that every multiassociative search binds the participating representations through the particular Hebbian mechanisms originally proposed.
That implementation remains an empirical hypothesis.
12. Iterative Compounding and Multistep Reasoning
The model is particularly relevant to reasoning because complex reasoning usually requires intermediate products that cannot be discarded immediately after they are computed.
Consider:
\{A,B,C\}\rightarrow D
followed by:
\{A,C,D\}\rightarrow E
and then:
\{A,D,E\}\rightarrow F.
The result D is not a final answer. It becomes an input to a subsequent computation. The same is true of E.
This process is iterative compounding.
Research on relational integration provides evidence that higher-order cognition requires multiple pieces of information to be combined rather than maintained independently. Prabhakaran, Narayanan, Zhao, and Gabrieli (2000), for example, found increased prefrontal involvement when verbal and spatial information had to be integrated into a common representation.
Recent relational-integration research also indicates that representations of relational complexity emerge in higher-order cortical systems after initial stimulus processing, consistent with an integration stage in which multiple constraints must be brought together. The deeper evidential limitation is that most existing tasks do not track several internally generated intermediate products across successive updates.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
That is exactly the kind of experiment required to move from component evidence to architectural evidence.
13. Hierarchies of Neural Timescales
The iterative updating model also proposes that relatively enduring higher-order representations interact with faster-changing sensory representations.
Independent neurophysiological evidence supports such temporal heterogeneity.
Murray and colleagues (2014) identified a hierarchy of intrinsic neural timescales across primate cortex. Sensory regions generally exhibited relatively short temporal integration windows, whereas association regions, including prefrontal cortex, exhibited longer intrinsic timescales.
Geerligs and colleagues (2022) likewise found partially nested cortical states during naturalistic perception, with comparatively shorter state durations in sensory regions and longer states in higher-order regions.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
Such findings are compatible with a system in which enduring abstract constraints shape more rapidly changing sensory or imagery states.
However, three variables should remain distinct:
\text{intrinsic neural autocorrelation timescale}
\neq
\text{lifetime of a represented concept}
\neq
\text{icSSC half-life}.
The existence of a cortical timescale hierarchy establishes biological plausibility, not the specific temporal quantities proposed by the model.
Furthermore, persistence should not automatically be equated with cognitive superiority. Lugtmeijer and colleagues found that neural-state durations can lengthen with aging in several regions, illustrating that longer states are not intrinsically advantageous.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
The stronger prediction is adaptive persistence: relevant information should remain available as long as it continues to constrain processing, whereas irrelevant information should be released efficiently.
14. Relationship to Retrieved-Context Models
The closest theoretical relatives of iterative updating are not theories in which memory is simply reset after every operation.
Howard and Kahana’s (2002) Temporal Context Model proposes a continuously evolving context that influences retrieval. Retrieval reinstates aspects of earlier context, thereby modifying the contextual state that controls subsequent memory search.
Polyn, Norman, and Kahana’s (2009) Context Maintenance and Retrieval model extends this principle by allowing several kinds of contextual information to organize free recall. Later CMR formulations model additional properties of retrieval over longer temporal spans.
The structural similarity is substantial.
Retrieved-context models can be summarized as:
\text{retrieve item}
\rightarrow
\text{update context}
\rightarrow
\text{updated context retrieves next item}.
Iterative updating proposes:
\text{generate item}
\rightarrow
\text{item joins retained representational set}
\rightarrow
\text{updated set generates next item}.
The distinction should therefore not be framed as iteration versus noniteration. Retrieved-context theories already employ iterative contextual dynamics.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
The potentially discriminating difference concerns representation.
In TCM and CMR, the effective retrieval cue is a distributed contextual state shaped by history and reinstatement.
In the iterative updating model, the operative state is more explicitly characterized as a collection of selectively retained, semantically interpretable representations whose individual members can persist, disappear, reenter, and change relative influence.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
That difference can be tested only if the identities of the retained representations are measured independently.
15. Relationship to Working-Memory Gating Models
Selective updating itself is not novel.
Prefrontal-basal-ganglia models such as that of O’Reilly and Frank (2006) already formalize learned gating into and out of working memory. Such mechanisms can explain why one representation remains maintained while another is replaced.
This suggests a complementary relationship.
Gating mechanisms answer:
What should remain available?
Multiassociative iteration asks:
What does the retained configuration do once it remains available?
The distinctive hypothesis is therefore not the existence of selective gating but the repeated use of selectively retained contents in generative computation.
This refinement significantly strengthens the novelty claim.
Rather than arguing that previous models failed to consider updating or recurrence, the theory can propose that several known mechanisms participate in a common cycle:
\text{selective retention}
\rightarrow
\text{joint constraint}
\rightarrow
\text{successor generation}
\rightarrow
\text{reweighting}
\rightarrow
\text{renewed search}.
The empirical review supports replacing broad claims that psychology or neuroscience ignored iteration with this more specific architectural proposal.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
16. Four Levels of Evidence
The empirical status of the model becomes clearer if findings are divided into four categories.
16.1 Component evidence
A single required operation is established.
Examples include persistent content-sensitive activity, item-specific removal, latent memory accessibility, multiple-prime facilitation, mental transformation, and temporal bridging.
This category is now extensive.
16.2 Coupled-mechanism evidence
Two or more operations interact in the same experiment.
Sawagashira and Tanaka combine persistence, relevance-sensitive removal, content-specific neural signals, and perturbational evidence.
Warden and Miller show retention together with representational modification following new input.
Rose and colleagues link latent retention, reactivation, and behavioral consequences.
The model has substantial evidence at this level.
16.3 Architectural evidence
Retention, endogenous successor generation, and renewed processing are linked over successive updates.
Compound cuing and context-reinstatement experiments approach this level because prior internal state predicts self-generated retrieval.
Yet most do not independently identify the individual representations that constitute the operative state.
Architectural evidence is therefore suggestive but incomplete.
16.4 Discriminating evidence
A distinctive prediction of iterative updating outperforms credible alternatives.
This is the weakest current category.
The most informative comparison is likely to involve:
- newest-item association,
- recency-weighted history,
- retrieved temporal context,
- selectively retained representational contents.
The theory will become much stronger when these models are tested against the same behavioral and neural data.
17. A More Precise Empirical Formulation of the Theory
The accumulated evidence supports a narrower and more defensible formulation than some of the broader statements in earlier versions of the model:
Human cognition frequently proceeds through selective, partial updating of a limited set of task-relevant representations. Representations retained across transitions contribute to the interpretation and selection of newly generated contents. Newly generated contents alter the representational configuration and thereby change the conditions governing subsequent retrieval, imagery, inference, or action. Repetition of this retain-add-reweight process permits cumulative cognitive trajectories in which intermediate products remain available long enough to constrain later products.
This formulation emerged from the evidence synthesis and captures the portion of the theory for which every major operation presently has empirical support.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
It is also experimentally tractable.
The remaining burden is not demonstrating that memory persists, that updating occurs, or that context affects retrieval. Those propositions are established.
The remaining burden is demonstrating their repeated causal coordination.
18. Three Discriminating Experiments
18.1 Behavioral experiment: relevance against recency
A behavioral experiment should explicitly dissociate how recently information appeared from whether it remains relevant.
Participants could perform a sequence of constrained associative or relational problems. Four meaningful representations would initially be supplied. During subsequent operations, one relatively old representation would remain essential while a more recent representation would become obsolete.
For example:
S_1=\{A,B,C,D\}
S_2=\{A,C,D,E\}
S_3=\{A,C,E,F\}.
Here, A remains relevant throughout, while other representations enter and leave.
Participants would then generate G, rather than selecting it from supplied alternatives.
Four models could be compared:
M_1:G\sim\text{latest item}
M_2:G\sim\text{recency-weighted history}
M_3:G\sim\text{retrieved contextual state}
M_4:G\sim\text{selectively retained representational set}.
Strong support would occur if older but still relevant A predicts G more strongly than newer but obsolete information after association strength, recency, rehearsal, and fixed task instructions are controlled.
The decisive second step is:
\{A,C,E,G\}\rightarrow H.
If G then contributes to generating H, the experiment begins to test iterative compounding rather than selective memory alone.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
18.2 Neural experiment: reconstructing the evolving state
Participants could first complete a content-localizer task establishing neural signatures for objects, concepts, relations, goals, or rules.
They would then undertake a multistep problem in which several constraints must be maintained while intermediate results are generated internally.
The predicted neural state might resemble:
Time
A
B
C
D
E
F
t_1
high
high
high
high
low
low
t_2
high
low
high
high
high
low
t_3
high
low
high
low
high
high
This is the representational equivalent of the staggered activity pattern proposed by the model.
The crucial analysis is not whether successive whole-brain states resemble one another.
It is whether neural evidence for retained A,C,D at t_1 predicts the identity of endogenous E at t_2, and whether the emergence of E then improves prediction of F.
Intracranial recording would provide high temporal resolution when clinically feasible. MEG or EEG decoding could provide a scalable human approach. fMRI would provide useful anatomical information but would require careful control for hemodynamic temporal smoothing.
The target relationship is:
S_t\rightarrow x_{t+1}
followed by:
(S_t-D_t+x_{t+1})\rightarrow x_{t+2}.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
18.3 Causal experiment: perturbing a retained constraint
The strongest experiment would identify a specific retained representation, show that it predicts the next cognitive product, and then perturb it.
Suppose representation A has been present longer than a competing representation but remains necessary for generating G.
The experimental sequence would be:
\text{identify retained }A
\Downarrow
A\text{ predicts }G
\Downarrow
\text{perturb }A
\Downarrow
G\text{ changes specifically}.
The prediction is not generalized slowing or increased error.
It is a content-specific shift in successor selection.
If the resulting change in G subsequently changes H, the experiment would establish:
A\rightarrow G\rightarrow H
within a recurrently updated state.
This would approach a smoking-gun demonstration of the proposed architecture.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
19. Implications for the Study of Thought
The empirical question traditionally posed in working-memory research is often:
How is information retained?
The iterative framework adds a second:
How is retained information used to produce the next state?
This change of emphasis is consequential.
Thought is not simply memory plus processing. A train of thought consists of processing whose products alter the conditions of subsequent processing.
A state can therefore function simultaneously as:
- the outcome of preceding computation,
- temporary memory for intermediate information,
- the contextual interpretation of current information,
- a search configuration for the next cognitive product.
This recursive causal organization explains why intermediate products can accumulate rather than disappear.
An inference can become a premise.
A remembered feature can become a constraint.
A feature discovered in imagery can become a conceptual variable.
A provisional action can expose new information that changes the next plan.
A retrieved memory can alter the context from which another memory is retrieved.
Iterative updating provides a common computational description of these superficially different processes.
20. Implications for Consciousness
The present evidence bears most directly on cognition, not phenomenal consciousness.
Working-memory maintenance, contextual retrieval, neural persistence, imagery transformation, relational integration, and associative learning can be studied behaviorally and physiologically. None of the findings reviewed here establishes that iterative updating is sufficient for subjective experience.
The 2016 formulation itself acknowledged that the model was qualitative, exploratory, and contained untested assumptions.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
A more rigorous research program should therefore separate two hypotheses:
\text{iterative updating}
\rightarrow
\text{cognitive continuity}
from:
\text{cognitive continuity}
\rightarrow
\text{phenomenal continuity}.
The first can increasingly be tested using objective measurements.
Only after its cognitive architecture has been established should the second question be used to evaluate whether the same mechanisms contribute to consciousness.
This separation does not weaken the consciousness hypothesis. It makes it scientifically tractable by preventing evidence for working memory or reasoning from being counted prematurely as evidence for phenomenal experience.
21. Implications for Artificial Cognitive Architectures
The framework also suggests a computational design principle.
An artificial system capable of extended cognition should not merely retain a transcript of previous processing. It should maintain a selectively prioritized working state in which earlier intermediate products remain available only while useful, interact with newly generated information, and influence later generation.
The relevant algorithm is approximately:
- Maintain a limited set of active or rapidly recoverable representations.
- Estimate the continued relevance of each representation.
- Remove or demote contents whose expected utility has declined.
- Use the retained configuration jointly to generate candidate additions.
- Incorporate the selected addition.
- Reweight the retained contents in light of the addition.
- Repeat.
This differs from simply extending a context window. A large passive history contains information, but iterative updating requires active state management and recursive generative reuse.
Such an architecture may be especially valuable for long-horizon reasoning, planning, multimodal simulation, and internally generated learning because intermediate products can remain operative without requiring the system to reconstruct its entire previous computation at every step.
22. Discussion
The empirical literature reviewed here produces a consistent pattern.
First, partial updating is real. Working memory can preserve selected representations while removing others.
Second, retention is relevance-sensitive. Neural persistence need not obey a simple recency schedule.
Third, retained information can be represented in several physiological formats. Sustained firing is important but is not the only plausible mechanism.
Fourth, representations are context-sensitive. Newly arriving information can alter the coding and behavioral influence of information already in working memory.
Fifth, several sources of information can jointly influence subsequent processing.
Sixth, internal context predicts self-generated retrieval.
Seventh, imagery can be actively transformed and can generate novel information.
Eighth, maintained activity can link temporally separated events and influence future learning.
Together these findings support:
\text{retention}
\rightarrow
\text{integration}
\rightarrow
\text{selection}
\rightarrow
\text{updating}
\rightarrow
\text{further processing}.
The empirical review accordingly concludes that features concerning selective maintenance and partial updating receive particularly strong support, while the hardest part of the architecture lies in features 5 through 7, where retained representations are proposed to generate successive endogenous additions repeatedly.
Empirical Evidence for the Iterative Updating Model of Cognition.pdf
The critical absence in the literature is therefore highly specific.
It is not the absence of evidence for persistence.
It is not the absence of evidence for partial updating.
It is not the absence of evidence for contextual retrieval.
It is not the absence of evidence that several cues can converge.
It is the absence of a direct experiment demonstrating:
\boxed{
\text{retained identifiable contents}
\rightarrow
\text{endogenous successor}
\rightarrow
\text{updated identifiable contents}
\rightarrow
\text{next endogenous successor}
}
with causal influence demonstrated at each transition.
That evidential gap is scientifically useful because it converts a broad cognitive proposal into a sharply testable hypothesis.
23. Conclusion
The evidence now supports a substantial portion of the iterative updating model.
Working memory can be selectively and partially updated. Representations can persist for different intervals according to relevance. Retained contents can occupy active or latent neural states. New information can transform the effective representation of existing information. Multiple cues and reinstated contexts can influence later retrieval. Internally generated representations can be transformed through mental imagery. Neural activity can bridge events separated in time, and working-memory processing can alter long-term memory.
These findings collectively establish the plausibility of a generative cognitive architecture in which a changing working-memory state carries selected information forward and uses it to shape subsequent computation.
The principal claim still requiring direct demonstration is the repeated causal coordination of these mechanisms during endogenous thought.
If an experiment can show that a set of independently identified retained representations predicts a newly generated cognitive item, that the new item joins and alters the retained configuration, and that this altered configuration predicts another internally generated item, then the central architectural hypothesis will have been tested directly.
The resulting principle is simple:
\boxed{
S_t\rightarrow x_{t+1}\rightarrow S_{t+1}\rightarrow x_{t+2}
}
but its consequences are extensive.
A thought need not be treated as an isolated event followed by another thought. Each cognitive state can preserve selected products of previous processing, add a new product, and thereby become the computational starting point for what follows.
On this view, the continuity and productivity of thought arise from the same operation: the selective preservation of information long enough for the products of one cognitive state to participate in constructing the next.
References
Balota, D. A., & Paul, S. T. (1996). Summation of activation: Evidence from multiple primes that converge and diverge within semantic memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 827-845. https://doi.org/10.1037/0278-7393.22.4.827
Christophel, T. B., Cichy, R. M., Hebart, M. N., & Haynes, J.-D. (2015). Parietal and early visual cortices encode working memory content across mental transformations. NeuroImage. https://doi.org/10.1016/j.neuroimage.2014.11.018
Dijkstra, N., et al. (2017). Distinct top-down and bottom-up brain connectivity during visual perception and imagery. Scientific Reports. https://doi.org/10.1038/s41598-017-05888-8
Ecker, U. K. H., Oberauer, K., & Lewandowsky, S. (2014). Working memory updating involves item-specific removal. Journal of Memory and Language, 74, 1-15. https://doi.org/10.1016/j.jml.2014.03.006
Finke, R. A., Pinker, S., & Farah, M. J. (1989). Reinterpreting visual patterns in mental imagery. Cognitive Science, 13, 51-78. https://doi.org/10.1207/s15516709cog1301_2
Gilmartin, M. R., Miyawaki, H., Helmstetter, F. J., & Diba, K. (2013). Prefrontal activity links nonoverlapping events in memory. Journal of Neuroscience. https://doi.org/10.1523/JNEUROSCI.0144-13.2013
Howard, M. W., & Kahana, M. J. (2002). A distributed representation of temporal context. Journal of Mathematical Psychology, 46, 269-299. https://doi.org/10.1006/jmps.2001.1388
Kamiński, J., et al. (2017). Persistently active neurons in human medial frontal and medial temporal lobe support working memory. Nature Neuroscience. https://doi.org/10.1038/nn.4509
Lavigne, F., Longrée, A., Mayaffre, D., & Mellet, E. (2016). Semantic integration by pattern priming: Experiment and cortical network model. Cognitive Neurodynamics. https://doi.org/10.1007/s11571-016-9410-4
Lewis-Peacock, J. A., Drysdale, A. T., Oberauer, K., & Postle, B. R. (2012). Neural evidence for a distinction between short-term memory and the focus of attention. Journal of Cognitive Neuroscience, 24, 61-79. https://doi.org/10.1162/jocn_a_00140
Lohnas, L. J., & Kahana, M. J. (2014). Compound cuing in free recall. Journal of Experimental Psychology: Learning, Memory, and Cognition, 40, 12-24. https://doi.org/10.1037/a0033698
Lohnas, L. J., Polyn, S. M., & Kahana, M. J. (2015). Expanding the scope of memory search: Modeling intralist and interlist effects in free recall. Psychological Review. https://doi.org/10.1037/a0039036
Manning, J. R., Polyn, S. M., Baltuch, G. H., Litt, B., & Kahana, M. J. (2011). Oscillatory patterns in temporal lobe reveal context reinstatement during memory search. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.1015174108
Manning, J. R., Sperling, M. R., Sharan, A., Rosenberg, E. A., & Kahana, M. J. (2012). Spontaneously reactivated patterns in frontal and temporal lobe predict semantic clustering during memory search. Journal of Neuroscience, 32. https://doi.org/10.1523/JNEUROSCI.5321-11.2012
Morton, N. W., & Polyn, S. M. (2016). A predictive framework for evaluating models of semantic organization in free recall. Journal of Memory and Language, 86, 119-140. https://doi.org/10.1016/j.jml.2015.10.002
Murray, J. D., et al. (2014). A hierarchy of intrinsic timescales across primate cortex. Nature Neuroscience. https://doi.org/10.1038/nn.3862
Murray, J. D., et al. (2017). Stable population coding for working memory coexists with heterogeneous neural dynamics in prefrontal cortex. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.1619449114
O’Reilly, R. C., & Frank, M. J. (2006). Making working memory work: A computational model of learning in the prefrontal cortex and basal ganglia. Neural Computation. https://doi.org/10.1162/089976606775093909
Panichello, M. F., & Buschman, T. J. (2021). Shared mechanisms underlie the control of working memory and attention. Nature, 592, 601-605. https://doi.org/10.1038/s41586-021-03390-w
Polyn, S. M., Norman, K. A., & Kahana, M. J. (2009). A context maintenance and retrieval model of organizational processes in free recall. Psychological Review. https://doi.org/10.1037/a0014420
Prabhakaran, V., Narayanan, K., Zhao, Z., & Gabrieli, J. D. E. (2000). Integration of diverse information in working memory within the frontal lobe. Nature Neuroscience, 3, 85-90.
Ranganath, C., Cohen, M. X., & Brozinsky, C. J. (2005). Working memory maintenance contributes to long-term memory formation: Neural and behavioral evidence. Journal of Cognitive Neuroscience. https://doi.org/10.1162/0898929054475118
Reser, J. E. (2016). Incremental change in the set of coactive cortical assemblies enables mental continuity. Physiology & Behavior, 167, 222-237. https://doi.org/10.1016/j.physbeh.2016.09.019
Reser, J. E. (2022-2026). A cognitive architecture for machine consciousness and artificial superintelligence: Thought is structured by the iterative updating of working memory. Manuscript and arXiv:2203.17255. The present article uses the attached revised manuscript as the theoretical source. Its abstract characterizes successive working-memory states as overlapping configurations whose retained contents participate in associative search.
Thought iterative updating(20260907-214548).pdf
Rose, N. S., et al. (2016). Reactivation of latent working memories with transcranial magnetic stimulation. Science, 354, 1136-1139. https://doi.org/10.1126/science.aah7011
Sabo, T., & Schneider, D. (2025). Processing in working memory boosts long-term memory representations and their retrieval. Communications Psychology. https://doi.org/10.1038/s44271-025-00309-3
Sawagashira, R., & Tanaka, M. (2025). Neural correlates of memory updating in the primate prefrontal cortex. Communications Biology, 8, 900. https://doi.org/10.1038/s42003-025-08271-w
Schlegel, A., et al. (2013). Network structure and dynamics of the mental workspace. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.1311149110
Taylor, R., Tomić, I., Aagten-Murphy, D., & Bays, P. M. (2023). Working memory is updated by reallocation of resources from obsolete to new items. Attention, Perception, & Psychophysics, 85, 1437-1451. https://doi.org/10.3758/s13414-022-02584-2
Warden, M. R., & Miller, E. K. (2007). The representation of multiple objects in prefrontal neuronal delay activity. Cerebral Cortex, 17, i41-i50. https://doi.org/10.1093/cercor/bhm070
Wolff, M. J., Jochim, J., Akyürek, E. G., & Stokes, M. G. (2017). Dynamic hidden states underlying working-memory-guided behavior. Nature Neuroscience, 20, 864-871. https://doi.org/10.1038/nn.4546
