A Working-Memory Mechanism for Creativity, Innovation, and the Progressive Transformation of Thought

Abstract

This article proposes that thinking is continuous iterative query construction. On this account, the active contents of working memory function as a composite query: a temporary coalition of perceptual, semantic, affective, motoric, mnemonic, and goal-related representations that jointly constrain the search for the next cognitive update. Thought proceeds when multiassociative search selects an update, some prior contents are retained, others are removed or inhibited, and the updated state becomes the next query. This process allows thought to preserve continuity while progressively transforming its contents.

The proposal builds on the iterative updating model of working memory, according to which successive states of working memory overlap because some representations remain active or potentiated across state transitions. This article develops a specific implication of that model: cognition is not merely search over memory or problem space, but the continuous reconstruction of the search state itself. Simple recall may involve familiar cues retrieving familiar completions. Reasoning involves constrained sequences of query revision. Creativity emerges when familiar representations form unfamiliar coalitions, generate intermediate products, and incorporate those products into subsequent query states. Originality therefore arises through successive rounds of constrained recombination, not through a separate faculty detached from ordinary cognition.

The article situates this mechanism in relation to working-memory theory, global workspace theory, associationism, predictive processing, conceptual blending, cognitive architectures, and artificial intelligence. It argues that these models capture important components of cognition, while continuous iterative query construction offers a candidate transition rule for how those components operate as a moving stream of thought. The article concludes with predictions concerning representational overlap, insight, fixation, expertise, distraction, and artificial systems capable of autonomous discovery.

Keywords

continuous iterative query construction; iterative updating; working memory; multiassociative search; creativity; innovation; thought; reasoning; mental continuity; associative search; cognitive architecture; focus of attention; global workspace; conceptual blending; predictive processing; artificial intelligence; machine consciousness; superintelligence; scientific discovery; constrained recombination

1. Introduction: The Missing Transition Rule

Cognition is often described in terms of attention, memory, prediction, association, control, search, and representation. These are not mistaken constructs. Each names something real about the way minds organize information. Working memory temporarily maintains task-relevant contents. Attention selects and prioritizes them. Long-term memory preserves what has been learned. Associative mechanisms link related contents. Predictive mechanisms anticipate what is likely to happen next. Executive systems regulate behavior in accordance with goals. Global workspace models explain how selected contents become broadly available to specialized systems. Cognitive architectures specify buffers, operators, rules, utilities, goals, and procedures. Contemporary artificial intelligence systems demonstrate the power of attention, context, and learned statistical structure.

Yet a problem remains. These accounts often describe components of cognition more clearly than they describe the transition by which one thought becomes the next. The mind does not merely contain memories, goals, perceptions, and associations. It moves. It advances from one representational state to another while preserving enough continuity to remain on topic and changing enough content to make progress. A theory of thought must therefore explain not only what cognition contains, but how those contents are continuously transformed.

The present article proposes that thinking is continuous iterative query construction. On this account, the active contents of working memory function as a composite query. This query is not a linguistic prompt or a discrete search string, but a temporary coalition of active and recently potentiated representations. These representations jointly constrain the search for an update. When a new representation is retrieved, inferred, predicted, generated, or otherwise selected, it does not merely appear as an output. It becomes part of the next working-memory state. Because some previous contents are retained while others are removed or inhibited, the next state is a modified version of the previous query. Thought proceeds as this process repeats.

This framing builds on the iterative updating model of working memory, according to which successive states of working memory overlap in content. Some representations persist across state transitions, some lose activity, and others enter the active set. In this way, each working-memory state is a revised iteration of the preceding state. The retained contents preserve context, while the newly added contents transform the direction of subsequent processing. The result is not a sequence of isolated mental snapshots, but a continuous chain of partially overlapping states.

The term query is useful here because the contents of working memory are not inert. They do not merely wait to be inspected by a central executive. Rather, they collectively bias the next moment of cognition. A person thinking about a problem, image, hypothesis, memory, social exchange, or future action is holding a set of mutually constraining representations in a heightened state of availability. That set activates related contents in long-term memory, perceptual systems, semantic networks, motor representations, affective systems, and procedural models. The next thought is selected from this global field of possible updates.

However, the mind does not search with the same query over and over. It changes the query as it searches. Each update modifies the active state that will be used to select the next update. This is the central point. Thinking is not merely a search process. Thinking is a process that continuously constructs the search state itself.

This distinction helps separate several related cognitive operations. In simple recall, a familiar cue may retrieve a familiar completion. In ordinary pattern completion, partial information fills in a known structure. In associative thought, one idea may suggest another. In reasoning, a sequence of states may progress according to constraints learned from prior experience. In creative thought, however, familiar elements may be combined into an unfamiliar working-memory coalition. That coalition may produce an intermediate result that has not previously been stored as a completed pattern. Once incorporated into working memory, the result changes the next query, which changes the next search, which may produce another intermediate result. Originality may arise from such successive rounds of constrained recombination.

The contribution of this article is therefore not the claim that cognition involves memory, attention, association, prediction, or search. Nor is it the claim that previous models are wrong. Many existing models capture real and necessary elements of cognition. The proposed contribution is more specific: thought may be generated by a continuous transition rule in which working-memory states function as self-modifying composite queries. Each state searches from the standpoint of its current contents, incorporates the product of that search, and thereby becomes a new query state.

This hypothesis has implications for creativity and innovation. It suggests that original thought is not a separate magical faculty added onto ordinary cognition. Rather, creativity may emerge when iterative query construction carries familiar representations into unfamiliar combinations, preserves intermediate products long enough for them to compound, and constrains the resulting sequence by relevance, coherence, evidence, goals, or utility. Innovation, in turn, may occur when such internally generated products survive external testing and become useful to others.

This account also has implications for artificial intelligence and superintelligence. A system that merely answers prompts may perform impressive pattern completion. A system that can preserve unresolved problems, construct composite query states, revise those states with intermediate products, and continue searching across long chains of partially overlapping representations would be closer to an autonomous thinker. Such a system would not only search better. It would construct better and better queries.

The present article develops this proposal in stages. First, it situates continuous iterative query construction in relation to existing cognitive models. Second, it treats working memory as a composite query rather than a passive store. Third, it describes how iterative updating and multiassociative search can transform one query state into the next. Fourth, it applies this mechanism to recall, reasoning, creativity, and innovation. Finally, it outlines predictions and possible tests. The central thesis is simple: thought progresses because the mind continuously revises the query from which the next thought is selected.

2. Existing Models Capture Real Components of Cognition

The proposal advanced here is intended as an integrative account, not as a rejection of prior models. Cognitive science, neuroscience, psychology, artificial intelligence, and philosophy have produced many frameworks that capture important aspects of mind. A satisfactory theory of thought should not discard these models. It should explain how their valid components can operate together across time.

Classical memory models established that cognition depends on interactions between sensory memory, short-term or working memory, and long-term memory. The multi-store tradition distinguished brief sensory traces, temporary active maintenance, and durable storage. Later working-memory theories refined this picture by emphasizing limited capacity, domain-specific buffers, executive control, rehearsal, and the manipulation of information in the service of ongoing cognition. Baddeley and Hitch’s multicomponent model, for example, helped identify separable verbal, visuospatial, episodic, and control-related aspects of working memory. Cowan’s embedded-processes model sharpened the idea that working memory is not a separate container into which memories are copied, but an activated subset of long-term memory containing a still more active focus of attention.

These views provide essential foundations. They explain why thought requires temporary availability, attentional priority, limited capacity, and interactions between active and inactive memory. However, they often leave open a deeper dynamical question: how does the current active set become the next active set? Working memory is known to update, but updating is often treated as a task operation, a capacity-limited manipulation, or a function attributed to a central executive. The transition rule itself remains underspecified. Continuous iterative query construction attempts to specify this rule by treating each working-memory state as a search condition that is modified by its own products.

Global Workspace Theory and related blackboard architectures provide another crucial component. They explain how selected information becomes globally available to otherwise specialized and partially isolated systems. A content that enters the global workspace can influence perception, memory, planning, verbal report, motor preparation, and evaluation. This helps explain why conscious contents have a broad causal reach. It also fits naturally with the present proposal, because a composite query must be globally available if it is to search many memory and processing systems at once.

Still, global availability is not the same thing as continuous transformation. A workspace can broadcast content without fully explaining how one broadcast state becomes the next. If the global workspace is treated only as a stage on which selected contents appear, then another process must still determine how the current cast of contents gives rise to the next cast. Continuous iterative query construction proposes that the contents of the workspace jointly participate in selecting their successors. The workspace is not merely a display. It is part of the search state.

Associationist theories also capture something fundamental. Mental contents do call up related contents. Similarity, contiguity, contrast, cause and effect, means and end, premise and conclusion, and other associative relations clearly influence the stream of thought. However, a simple one-association-at-a-time view is too narrow. Thought is rarely just A evoking B. More often, several active contents jointly constrain what comes next. A remembered person, a current goal, an emotional state, a physical setting, a linguistic phrase, and a latent problem can all contribute to the selection of the next thought. The present model therefore extends associationism from a linear relation between successive ideas to a coalition-based process in which the active contents of working memory jointly search for the next update.

Predictive-processing and related Bayesian models also capture an indispensable aspect of cognition. The brain is constantly anticipating what is likely, explaining away uncertainty, and updating internal models in light of error, surprise, or new information. This is highly compatible with iterative query construction. A working-memory state can be understood as a temporary model of the current situation, and the next update can be understood as a prediction, correction, or inferred addition. However, prediction alone does not fully specify how the active contents of working memory are retained, replaced, recombined, and transformed into the next search condition. Iterative query construction focuses on this transformation.

Conceptual blending and related theories of analogy, metaphor, and mental-space integration also provide a powerful account of creativity. They show how distinct domains can be selectively projected into a blended space, producing emergent structure not present in either input alone. This is central to human originality. The present proposal does not reject blending. It attempts to place blending inside a temporal working-memory mechanism. A blend may be one product of iterative query construction. It may also become a new query element that enables further search. From this perspective, creativity does not consist only of constructing a blend. It also consists of using the blend as a new search state.

Cognitive architectures such as ACT-R, Soar, LIDA, OpenCog, and related systems have contributed detailed models of goals, buffers, operators, production rules, declarative memory, procedural memory, reinforcement, attention, and action selection. These architectures are valuable because they force cognitive theory to become operational. They show how multiple components can be organized into systems that perform tasks. However, many such architectures rely heavily on symbolic rules, buffers, utilities, operators, or executive procedures to determine state transitions. The present proposal is different in emphasis. It suggests that the production sequence of thought may arise from subsymbolic coalition dynamics within working memory: active representations jointly search, one or more updates are selected, and the updated coalition becomes the next query.

Contemporary artificial intelligence adds another important comparison. Transformer-based language models show the extraordinary power of attention over context. Given a context window, such systems can predict, complete, summarize, translate, reason, and generate language with remarkable flexibility. They demonstrate that the selection and weighting of contextual information can produce complex behavior. Yet the analogy to human thought remains incomplete. A context window is not necessarily a self-sustaining stream of thought. A model may use context to generate an output without preserving an autonomous working-memory state that continuously updates itself in the absence of external prompting. Iterative query construction points toward a system in which the active internal state is not merely used to produce an answer, but is recursively revised to produce the next internal state.

Search-based models of reasoning also capture a real aspect of cognition. Problem solving often involves exploring a space of possibilities, evaluating candidate paths, pruning unhelpful branches, and advancing toward a goal. But standard search metaphors can be misleading if the query is assumed to remain fixed. Human thought frequently changes the problem representation while trying to solve the problem. A partial answer changes what counts as relevant. An analogy changes the apparent structure of the task. A failed attempt changes the constraints. A new intermediate concept changes the next question. In creative cognition, the search space is not merely explored. It is reconstructed.

For this reason, the models reviewed above should be viewed as complementary rather than mutually exclusive. Working-memory theory identifies temporary active contents. Global workspace theory explains broad availability. Associationism explains linkage. Predictive processing explains anticipatory modeling. Conceptual blending explains emergent structure from integration. Cognitive architectures explain goals, procedures, and task organization. Attention-based AI demonstrates the power of context-sensitive weighting. Search models explain exploration through possibility spaces.

Continuous iterative query construction attempts to describe the process that allows these components to operate as a moving stream. At each moment, active and recently active representations form a composite query. This query recruits, predicts, or generates an update. The update modifies the query, while retained contents preserve continuity. The revised query then initiates the next moment of search. This mechanism may help explain how cognition remains coherent without becoming static, flexible without becoming chaotic, and creative without becoming random.

3. Working Memory as a Composite Query

The present proposal depends on treating working memory not merely as a store, buffer, or workspace, but as a dynamically constructed query. In ordinary language, a query is often understood as a verbal request, a search term, or an explicit question. That is not the intended meaning here. In this context, a query is the functional search condition created by the current state of the cognitive system. It is the set of active and recently potentiated representations that jointly determine what information will become most available next.

A working-memory state can therefore be understood as a composite query. It is composite because it contains multiple representational elements. These may include perceptual contents, linguistic tokens, remembered objects, emotional valuations, bodily states, motor intentions, goals, unresolved problems, imagined scenes, temporal markers, social meanings, and abstract concepts. It is a query because these elements are not inert. While active, they jointly bias the activation of related contents throughout long-term memory and across specialized neural systems.

This conception differs from a view in which working memory simply holds information for inspection by an executive controller. A person does not first place several items in working memory and then have a separate homuncular process decide what they mean. Rather, the items themselves participate in determining what will happen next. Their coactivation creates a field of constraint. The next update to thought is selected from the field of possibilities made available by the current coalition.

For example, the working-memory contents involved in thinking about a scientific problem may include a known empirical pattern, an unexplained anomaly, a candidate mechanism, an analogy from another domain, a methodological limitation, a remembered paper, and a present goal. None of these elements alone constitutes the whole thought. Together, however, they create a search state. They make some associations more probable, some predictions more accessible, some actions more likely, and some interpretations more coherent. The working-memory state functions as a multidimensional query over memory, perception, language, imagery, action, and evaluation.

The term composite query also helps clarify why thought is not reducible to linear association. If only one active representation determined the next representation, the stream of thought would be a chain of pairwise transitions. But most thought appears to depend on conjunctive constraint. A current topic, a goal, a perceptual context, a remembered fact, and an emotional concern can all contribute to what comes next. The next thought is not simply the strongest associate of any one item. It is the product of their combined relevance structure.

This is especially apparent in problem solving. A problem state is not a single cue. It is a configuration of constraints. When solving a puzzle, writing an argument, interpreting social behavior, or developing a hypothesis, the mind must maintain several interdependent conditions at once. These conditions define the space of acceptable next moves. The retained contents of working memory indicate what must remain true, what has already been tried, what remains unresolved, and what kind of update would count as progress. The working-memory state is therefore not merely the content of thought. It is the operative query from which the next thought is selected.

This view also explains why the same item can have different implications in different contexts. A representation does not contribute a fixed meaning independent of the coalition in which it appears. The concept “water,” for example, will participate differently in a thought about thirst, a chemical reaction, a flooded basement, a plant, an ocean, a memory, or a planetary biosignature. The local coalition determines which aspects of the representation become relevant. Thus, the composite query does not only select the next item. It also contextually weights the current items themselves.

This has important implications for the relation between symbolic and subsymbolic cognition. At the psychological level, working-memory items can appear to be symbolic: a word, a person, a goal, a rule, an object, a premise, or a conclusion. At the neural level, however, each item is implemented by distributed assemblies and weighted connections. The apparent symbol is supported by many subsymbolic components. When several such items are coactive, their subsymbolic components interact in parallel. The search conducted by working memory is therefore not only a search among discrete symbols. It is also a distributed search through the underlying statistical structure encoded in the network.

This helps explain how explicit reasoning can be underwritten by implicit processes. A person may experience a thought as a sequence of propositions, images, or intentions. But the transitions between those experienced contents may be selected by unconscious convergence among many distributed representations. The query is not necessarily something the thinker can verbally state. It is the total active condition of the system. Verbal questions are only one special case of a more general phenomenon.

The composite-query view also provides a bridge between working memory and long-term memory. Long-term memory supplies the stored structure through which the query operates. Working memory supplies the temporary activation pattern that determines which parts of that structure are searched. When items enter working memory, they do not leave long-term memory and move into a separate storage medium. Rather, long-term representations become active, attended, potentiated, or otherwise prioritized. The query is therefore embedded within memory itself. It is a temporary pattern of activation inside the same system that stores the associations, regularities, and abstractions being searched.

Because working memory is capacity-limited, the composition of the query is selective. Not all possible relevant information can be maintained at once. The system must prioritize. Some contents enter the focus of attention. Others remain in a more latent but still potentiated state. Still others decay, are suppressed, or return to baseline. This selective constraint is not merely a limitation. It may be essential to thought. A query that included everything would select nothing in particular. Intelligence depends on constructing a query that is narrow enough to constrain search but broad enough to permit useful discovery.

The short-term store and the focus of attention may contribute differently to this process. The focus of attention may hold the most active, reportable, and manipulable contents. The broader short-term store may preserve recently relevant material that no longer occupies focal awareness but continues to bias search. In this way, the active query may have a center and a periphery. The center determines the immediate object of thought. The periphery supplies continuity, latent context, and unresolved constraints. Together, they create a search condition richer than the narrow contents of focal awareness alone.

This distinction becomes important for extended thought. A person can temporarily shift attention to a subproblem while a larger objective remains active in the background. When the subproblem yields a useful result, that result can be merged back into the larger query. The mind can therefore decompose complex tasks into partial searches while preserving enough background context to reassemble their products. This capacity may be central to reasoning, planning, and creativity.

Working memory as a composite query also clarifies why thoughts are sensitive to recent history. The current state is not determined only by present sensory input or by a single retrieved memory. It is shaped by a trail of prior states. Recently active representations remain more available than baseline contents, either because they are still firing, potentiated, primed, or more easily reactivated. Thus, the current query contains residues of earlier queries. This allows a line of thought to maintain direction over time.

The central claim of this section can be stated simply: working memory is not only the place where thought occurs. Working memory is the query by which thought proceeds. Its active contents select, weight, and constrain the next update. When that update is incorporated into working memory, the query changes. The next section describes this transformation in terms of iterative updating and multiassociative search.

4. Iterative Updating and Multiassociative Search

If working memory functions as a composite query, then a theory of thought must explain how that query changes. The present article proposes that the query changes through iterative updating and multiassociative search. These two processes are complementary. Iterative updating describes how working-memory contents are retained, removed, and added across time. Multiassociative search describes how the current coalition of contents selects the additions that will update the next state.

Iterative updating begins from the observation that consecutive states of working memory are not wholly independent. The contents active at one moment do not disappear all at once and get replaced by an entirely new set. Some representations persist. Others decline in activity or are inhibited. Others enter the active set. As a result, the next state of working memory partially overlaps with the previous state. This partial overlap is the basis of continuity.

This is important because thought requires both stability and change. If too much content is replaced at every moment, thought becomes fragmented. The system loses the topic, goal, scene, or problem frame that gives the sequence coherence. If too little content is replaced, thought becomes static, repetitive, or perseverative. The system remains trapped in the same representational configuration. Intelligent thought requires an intermediate regime in which enough content persists to preserve context and enough content changes to permit progress.

In the iterative updating model, each state of working memory is a revised iteration of the preceding state. The retained items carry forward the context. The removed items reduce interference, mark completed operations, or reflect declining relevance. The added items introduce new information, predictions, associations, inferences, percepts, or candidate actions. Because the next state contains a subset of the prior state plus new content, it is both continuous with and different from what came before.

This transition can be described at multiple levels. At the neural level, persistent activity and short-term synaptic potentiation provide plausible mechanisms for temporary maintenance. Some neuronal assemblies remain active or potentiated across intervals, while others return to baseline and others become newly active. At the cognitive level, the thinker experiences this as a line of thought in which a topic or problem remains present while its details change. At the algorithmic level, the system updates a temporary state by partial replacement.

Multiassociative search specifies how new contents are selected. The term refers to a search process in which multiple coactive representations jointly contribute to the selection of the next update. The active coalition spreads activation through the network. Candidate representations receiving sufficient convergent support become more likely to enter working memory. Candidate representations receiving insufficient support remain inactive. Current representations that continue to receive support are retained. Current representations that lose support are demoted, inhibited, or forgotten.

This is different from simple association. In a simple associationist account, one idea evokes the next. In the present model, the active coalition as a whole searches for the next update. The relevant transition is not A leading to B. It is A plus B plus C jointly selecting D. Then B plus C plus D jointly select E. Then C plus D plus E jointly select F. The sequence is associative, but not merely pairwise. It is coalition-based, context-sensitive, and recursively updated.

The phrase multiassociative search is meant to capture this distributed convergence. Each active item contributes some portion of activation energy, inhibition, weighting, or contextual constraint. The next update is selected by the combined influence of the whole state. In this sense, the query is not a list of independent search terms. It is an interacting system of constraints. The strongest candidate update is the one that best satisfies the current coalition.

This mechanism helps explain how working memory can move through conceptual space without requiring a separate homunculus to choose the next thought. The current state itself biases the system toward its successor. Executive control may still influence this process through goals, inhibition, attention, reward, and task demands. But the production sequence of thought need not be attributed entirely to a mysterious central controller. The contents of working memory help determine their own replacement.

Multiassociative search also explains why the same representation may lead to different updates in different contexts. The item “bird” might activate “wing,” “song,” “migration,” “dinosaur,” “flu,” “cage,” “flight,” “feather,” or “symbol,” depending on the other contents currently active. The active coalition determines which associative pathway is most relevant. This makes thought flexible without making it arbitrary.

The iterative aspect becomes crucial when the search product is incorporated into the next query. Suppose a working-memory state contains a problem, an analogy, a constraint, and a partial hypothesis. Multiassociative search may produce a candidate mechanism. If that mechanism enters working memory, it changes the query. The next search is no longer being conducted from the original problem state. It is being conducted from a modified state that now includes the candidate mechanism. This new state may activate a prediction, a counterexample, a methodological test, or a more abstract principle. Each update changes what the system is searching for, even if no explicit verbal question is asked.

This is the transition from search to iterative query construction. A fixed search process takes a query and returns a result. Iterative query construction takes a query, returns a result, incorporates that result into the query, and searches again from the revised state. The search process therefore changes its own input conditions as it proceeds. The system is not merely moving through a pre-given space. It is transforming the space by transforming the query used to explore it.

This process can also explain the progressive character of reasoning. In a logical argument, each step depends on prior steps while introducing new constraints. In a mathematical proof, intermediate results become resources for later operations. In planning, each imagined action changes the anticipated situation from which the next action is considered. In mental imagery, new details are added while stable features of the imagined scene are preserved. In scientific theorizing, a tentative interpretation changes which observations seem relevant. In each case, the current state of working memory is updated in a way that both preserves and transforms the problem.

Importantly, the updates selected by multiassociative search are not necessarily conscious before they enter working memory. Much of the search may occur below the threshold of awareness. The thinker may experience only the selected update: a remembered word, a visual image, an intuition, a possible explanation, a next sentence, a motor impulse, or a sudden insight. The search that produced it may remain inaccessible. This helps explain why thoughts can seem to “come to mind” without deliberate construction, while still being shaped by the active context.

The model also accommodates deliberate control. A thinker can intentionally maintain a problem, suppress an obvious but unhelpful association, rehearse a goal, introduce a new analogy, or return to an earlier state. Such control does not replace multiassociative search. It changes the contents and weights of the query. Voluntary attention modifies which representations remain active, which are excluded, and which candidates are allowed to enter. Deliberation can therefore be understood as the guided manipulation of query composition.

This account also helps distinguish productive thought from distraction. In productive thought, updates remain relevant enough to preserve the thread while introducing useful change. In distraction, an update captures the query and redirects it away from the prior goal. In fixation, the system retains too much of the prior state and fails to incorporate useful alternatives. In confusion, the active coalition may contain incompatible or weakly related constraints that fail to converge on a stable update. In insight, a newly introduced representation reorganizes the relevance structure of the retained contents, allowing the coalition to converge on an update that had previously been inaccessible.

The most important implication is that the stream of thought is not merely a series of retrieved contents. It is a recursive sequence of partially overlapping search states. Each state is produced by the previous state and becomes the basis for producing the next. Thought is therefore continuous not because nothing changes, but because change is structured by retention. It is iterative not because the same operation repeats mechanically, but because each operation is applied to the result of the previous operation. It is query construction because the active state that selects the next update is itself under continuous revision.

Thus, iterative updating and multiassociative search together provide a candidate transition rule for cognition. Working memory forms a composite query. Multiassociative search selects an update. Iterative updating incorporates that update while preserving part of the prior state. The revised coalition becomes the next query. Through repetition, this cycle can produce recall, reasoning, imagery, planning, problem solving, creativity, and innovation. The next section develops this claim more explicitly by distinguishing search from continuous iterative query construction.

5. From Search to Continuous Iterative Query Construction

Search is one of the most familiar metaphors for cognition. A person searches memory for a name, searches a problem space for a solution, searches perceptual input for a target, searches language for the right word, or searches possible actions for the next move. This metaphor is useful because cognition often involves selection from a field of possibilities. However, the metaphor is incomplete if it treats the query as fixed.

A fixed-query search begins with a defined input condition and returns a result. The query is specified before the search begins. The search process may be efficient or inefficient, narrow or broad, shallow or deep, but the input condition remains relatively stable. This model is appropriate for many forms of retrieval. If a person is trying to remember the name of a familiar actor, a set of cues may converge on a stored answer. If a person sees the beginning of a familiar phrase, the rest of the phrase may be completed. If a person is asked a factual question, the answer may be retrieved from memory. In these cases, the query activates a preexisting pattern strongly enough for completion.

Many cognitive operations do involve this kind of pattern completion. The system receives partial information and reconstructs a familiar whole. Pattern completion is powerful because it allows a mind to use incomplete cues to recover stored structure. It is also essential for perception, language, memory, and action. A face partially occluded by shadow, a word heard in a noisy room, a melody recognized from a few notes, or a social situation inferred from a gesture all require the completion of patterns that have been learned through prior experience.

But thought is not exhausted by pattern completion. A mind can ask a question it has never asked before. It can place familiar concepts into an unfamiliar relation. It can construct a new analogy, generate a novel hypothesis, reinterpret a memory, invent a tool, formulate a theory, or discover a hidden constraint. In these cases, the system is not merely retrieving a completed pattern that was already stored. It is building a search state that may never have previously existed.

This is the transition from search to continuous iterative query construction. A mind does not simply search with a query. It constructs the query while searching. The current working-memory state selects or generates an update. That update enters working memory, changes the state, and thereby changes the next query. The process repeats. The query is therefore not a static input to cognition. It is an evolving state produced by cognition.

The word continuous is important. The process is not best understood as a clean sequence of separate operations: query, answer, query, answer. Rather, the query state is constantly being revised. Some components remain active across multiple moments. Others fade gradually. Others are actively suppressed. New components are introduced. Recently active representations continue to bias selection even after leaving focal awareness. The next query is not wholly new and not wholly identical to the previous one. It is a partially transformed continuation of it.

This gives thought its distinctive temporal structure. Thought is not a string of disconnected outputs. Nor is it a static representation held in place until an executive process replaces it. It is a self-modifying stream. The active coalition at one moment becomes the basis for constructing the active coalition at the next moment. The system keeps moving because each update changes the conditions under which the next update will be selected.

This also explains why intermediate products matter. In a fixed-query search, the result may terminate the process. In iterative query construction, the result often becomes a new search term. A remembered fact may become a premise. A visual image may become a clue. A metaphor may become a model. A partial hypothesis may become a new constraint. A failed solution may reveal the structure of the problem. A surprising association may redirect the search toward a more abstract relation. The product of one moment becomes part of the query for the next.

This recursive incorporation of intermediate products is central to reasoning. Consider the act of constructing an argument. The first claim changes which evidence becomes relevant. The evidence changes which objection must be addressed. The objection changes how the claim must be qualified. The qualification changes the next inference. The final argument is not retrieved from memory as a complete object. It is assembled through a chain of revised query states. Each step constrains the next.

The same is true of scientific thinking. A researcher may begin with an anomaly, a known mechanism, and a body of background knowledge. This coalition may produce a candidate explanation. Once held in working memory, that explanation changes the search. It suggests predictions, counterexamples, measurements, analogies, and possible mechanisms. These in turn modify the next query. Over time, the original problem is transformed. The researcher is no longer searching from the same state that initiated the inquiry. The inquiry has constructed a new cognitive landscape.

This process can also occur without deliberate verbal reasoning. An artist may begin with an image, mood, material, and formal constraint. A composer may hold a motif, harmonic expectation, rhythmic pattern, and emotional goal. A mathematician may hold a diagram, definition, conjecture, and prior failed attempt. In each case, the active state searches for an update. The update modifies the next state. Thought advances by changing the conditions from which it is advancing.

Continuous iterative query construction therefore provides a way to characterize the generativity of thought. A system that only completes familiar patterns can produce useful behavior, but it remains bounded by the direct retrieval of learned structures. A system that constructs new query states can use learned structures to explore combinations that were not directly given in experience. It can create a temporary configuration, search from it, incorporate the result, and search again. This allows cognition to generate trajectories that are constrained by memory but not limited to direct replay.

The process also provides a framework for distinguishing different forms of cognition. Recall can be understood as a familiar query returning a familiar result. Reasoning can be understood as a sequence of constrained query revisions. Imagination can be understood as query construction over internally generated perceptual and conceptual states. Planning can be understood as query construction over possible future action states. Creativity can be understood as query construction that produces useful novelty. Innovation can be understood as creativity that survives external evaluation and becomes transmissible.

The central claim is therefore not that the mind searches. That is already clear. The central claim is that thought continuously constructs the query through which search occurs. Working memory does not merely store the inputs to thought. It builds the search condition. Multiassociative search selects an update. Iterative updating incorporates that update into the next working-memory state. Through this cycle, the mind changes what it is searching for while it searches.

This account helps explain how thought can be both continuous and transformative. Continuity comes from retained contents. Transformation comes from newly incorporated contents. Direction comes from the constraints imposed by the active coalition. Flexibility comes from the fact that the coalition can be revised. Originality comes from the fact that a newly constructed coalition can search from a state that has never existed before.

6. Creativity as Constrained Recombination Across Time

Creativity is often described as the production of something novel and useful. This definition is broadly correct, but it does not explain the mechanism by which novelty becomes useful rather than random. A random generator can produce novelty. A hallucination can produce novelty. A dream can produce novelty. What distinguishes creative thought is not novelty alone, but novelty under constraint.

Continuous iterative query construction offers a mechanism for such constrained novelty. Creative thought begins with familiar representations, but combines them into unfamiliar coalitions. These coalitions search memory and conceptual space for possible updates. Some updates are irrelevant, incoherent, redundant, or misleading. Others preserve enough of the problem structure while introducing enough new information to change the next search. When an update is incorporated into working memory, it becomes a new query element. The next search is now constrained by both the original context and the intermediate product.

This process can be described as constrained recombination across time. Recombination supplies variation. Constraint supplies selection. Time allows intermediate products to compound. A single recombination may produce an interesting association, but extended creative thought often requires many rounds of revision. Each round preserves some prior structure, adds something new, and alters what can be found next.

This helps demystify originality. An original thought does not need to appear fully formed. It can emerge through a sequence of small transformations. At each step, the current coalition makes certain updates more probable. The selected update changes the coalition. The revised coalition makes a new set of updates available. Over many iterations, the system may arrive at a representation that was not stored in memory as a completed pattern. The result feels like an insight, invention, interpretation, or discovery because it was constructed through the process rather than simply retrieved.

The same account explains why creativity is deeply dependent on memory. Creativity is sometimes contrasted with memory, as if originality requires escaping what has been learned. But creative thought cannot operate without stored structure. Long-term memory supplies the concepts, images, facts, skills, analogies, emotions, procedures, and constraints that can be recombined. Working memory supplies the temporary coalition in which recombination occurs. Iterative updating supplies the sequence through which recombinations can accumulate. Creativity is therefore not the opposite of memory. Creativity is memory recursively recombined under constraint.

The constraints are many. Semantic constraints determine which concepts can coherently relate. Causal constraints determine which mechanisms could plausibly produce which effects. Logical constraints determine which conclusions follow from which premises. Perceptual constraints determine what can form a stable image or scene. Motor constraints determine what actions are possible. Social constraints determine what meanings will be understood by others. Aesthetic constraints determine what feels balanced, elegant, surprising, or complete. Goal constraints determine what counts as progress. Empirical constraints determine what survives contact with evidence.

In creative cognition, these constraints are not all explicitly represented at once. Many operate implicitly through the structure of long-term memory and the dynamics of activation. A scientist may not consciously list every constraint on a hypothesis. A musician may not consciously compute every constraint on a melody. A writer may not consciously state every constraint on a sentence. Yet the active query is shaped by these constraints. Candidate updates that fit the coalition better are more likely to be selected, retained, elaborated, and tested.

This provides a way to distinguish creative insight from mere association. An association becomes creative when it reorganizes the query in a productive way. A remote association that does not help the system move forward is only a distraction. A surprising association that reweights the relevance of retained contents, resolves a tension, opens a new path, or generates a testable prediction becomes an insight. The difference lies in how the update modifies the next search state.

For example, a scientist attempting to explain a biological phenomenon may hold in working memory an observation, a contradiction in the literature, a developmental mechanism, an ecological pressure, and an analogy from another species. Multiassociative search may produce a possible connection among these elements. If the connection is superficial, it will fade. If it changes which evidence seems relevant, which mechanism seems plausible, and which prediction should be checked next, it becomes a productive update. The creative act is not only the association itself. It is the way the association changes the following query.

This also explains why creative work often involves incubation. During incubation, the focal contents of a problem may leave awareness, but related representations may remain potentiated, primed, or more easily reactivated. The query does not remain fully active, but residues of it persist in the short-term store or in altered accessibility within long-term memory. Later, a new perception, memory, or thought may reactivate the unresolved coalition in a modified context. The solution may seem sudden, but it may depend on prior partial query construction.

The model also explains why expertise supports creativity. A novice has fewer structured representations to recombine and fewer constraints to guide selection. An expert has richer memory, more abstract categories, better predictions, more refined error signals, and more useful analogies. Expertise makes the search space more structured. It also allows the working-memory query to include compressed high-level representations that stand for large bodies of knowledge. This enables more powerful recombinations with fewer active items.

At the same time, expertise can inhibit creativity when it overconstrains the query. If prior patterns dominate too strongly, the system may repeatedly retrieve familiar completions. This produces fixation. The problem is not too little memory, but too much constraint from habitual query states. Creative thought often requires enough control to preserve the problem and enough flexibility to alter the query away from the most practiced associations. Novelty and constraint must remain in productive tension.

This tension can be understood in terms of retention and replacement. Too little retention fragments the creative process. Intermediate products vanish before they can compound. Too much retention freezes the process. The same constraints keep selecting the same updates. Creativity requires an intermediate regime in which the problem frame persists, but enough new material enters to reshape the search. The working-memory state must remain stable enough to mean something and plastic enough to become something else.

This account also helps explain why analogies are so powerful in creative thought. An analogy introduces a structured source domain into the current query. Once incorporated, it changes which relations are salient in the target domain. It does not simply add one new item. It reweights an entire pattern of possible updates. A good analogy changes the search space. It allows the thinker to ask a different question without abandoning the original problem.

Creative construction can therefore be represented as a sequence:

A current problem state activates a candidate update.
The update enters working memory.
The updated state changes the relevance of existing contents.
The revised coalition searches again.
A new intermediate product appears.
That product becomes part of the next query.
The sequence continues until a stable, useful, or externally testable product emerges.

This sequence applies across domains. In writing, a sentence changes the next sentence that can be written. In mathematics, a lemma changes the proof space. In engineering, a prototype changes the design problem. In art, a mark on the canvas changes the composition. In conversation, a response changes the social and semantic field. In science, a hypothesis changes what evidence must be sought. Creative thought is the temporally extended construction of increasingly informative query states.

The model also suggests that originality exists on a continuum. Personal originality occurs when the constructed query state produces something new to the thinker. Cultural originality occurs when the product is new to a community. Scientific or technological innovation occurs when the product is not only new but explanatory, predictive, useful, reproducible, or operationalizable. The same cognitive mechanism may underlie all three levels, but the standards of evaluation differ.

This distinction matters because internal creativity is not sufficient for innovation. An internally generated idea must be stabilized, articulated, tested, and transmitted. It must survive contact with evidence, other minds, and practical implementation. Continuous iterative query construction may generate the candidate product, but innovation requires further selection outside the individual mind. The creative sequence must become communicable and externally constrained.

The central claim of this section is that creativity is not a separate faculty added onto ordinary cognition. It is an intensified and extended use of the same update cycle that supports thought more generally. Familiar representations are combined into unfamiliar query states. Multiassociative search selects intermediate products. Iterative updating incorporates those products into new queries. Constraint prevents the process from becoming random. Evaluation determines which products survive.

Originality, on this view, comes from successive rounds of constrained recombination in which each intermediate product changes the next search space. The mind does not simply search for creative ideas. It constructs the conditions under which creative ideas can be found.

7. Innovation, Discovery, and Superintelligence

The preceding sections treated creativity as constrained recombination across time. This section extends the same mechanism to innovation and discovery. Creativity, innovation, and discovery are related, but they should not be treated as identical. Creativity refers to the generation of novel and useful internal products. Innovation refers to the stabilization, articulation, implementation, or transmission of those products in a form that can alter behavior, technology, science, art, or culture. Discovery refers to the generation of a representation that corresponds to a real structure, pattern, mechanism, or possibility that was previously unknown to the system, and perhaps unknown to any system.

Continuous iterative query construction may help explain all three. A creative thought begins as a sequence of internal query revisions. An innovation emerges when the products of that sequence are expressed, tested, refined, and made usable. A discovery occurs when one of those products successfully identifies a structure in the world. In each case, the mind does not simply retrieve an answer. It constructs a succession of search states that make the answer reachable.

This distinction is important because not all original thoughts are discoveries. A novel association may be personally meaningful but empirically false. A metaphor may be artistically powerful but scientifically invalid. A theory may be elegant but unsupported. Conversely, a discovery may arise from many small, unromantic acts of query refinement rather than from a sudden flash of genius. Continuous iterative query construction allows for both possibilities. It explains sudden insight as the visible culmination of prior state construction, and it explains slow discovery as the accumulation of many constrained intermediate updates.

Scientific discovery provides the clearest example. A scientist rarely begins with a fully specified question and retrieves a fully specified answer. More often, the question itself changes during investigation. An anomaly leads to a possible mechanism. The mechanism changes which data matter. The data alter the hypothesis. The hypothesis suggests a test. The test reframes the anomaly. The original query is transformed through the process of attempting to answer it. This is not a failure of scientific reasoning. It is the structure of scientific reasoning.

The same process occurs in technological invention. An engineer begins with a problem state, a constraint, a material, a function, a prior design, and a goal. Each prototype changes the design problem. Each failure reveals a hidden constraint. Each working subsystem becomes a new component in the search for the larger system. The invention is not simply found in a preexisting solution space. The solution space is progressively constructed by the sequence of intermediate products.

This account also applies to philosophical and theoretical work. A thinker may begin with an unresolved conceptual tension. An initial distinction clarifies part of the problem. That distinction creates a new vocabulary. The vocabulary allows previously separate ideas to be held together. Their coactivation produces a more abstract principle. The principle changes the interpretation of the original problem. In this way, the theory is not retrieved. It is grown through a sequence of partially overlapping query states.

Innovation therefore requires both cognitive and external stabilization. Internally, a candidate idea must persist long enough to be elaborated. It must survive interference from competing thoughts and be incorporated into later working-memory states. Externally, it must be written, drawn, formalized, built, measured, taught, or otherwise encoded outside the mind. Externalization allows a fragile internal product to become part of a larger cultural memory system. Once externalized, it can function as a query element for other minds.

This is why language, notation, diagrams, mathematics, code, and scientific instruments are so powerful. They do not merely communicate completed thoughts. They help construct new queries. A diagram allows visual relations to become active together. A mathematical notation compresses a complex relation into a manipulable item. A written paragraph preserves intermediate reasoning that would otherwise decay. A laboratory instrument reveals patterns that can enter working memory as new constraints. Cultural tools extend iterative query construction beyond the biological brain.

The same principle applies to collective cognition. A scientific community can be understood as a distributed system that preserves, recombines, and updates query states across many individuals and artifacts. One person’s hypothesis becomes another person’s evidence. One laboratory’s anomaly becomes another laboratory’s problem. One theory becomes the background assumption for a new generation of questions. Culture allows query construction to become transindividual. The state is no longer maintained only in one brain, but across papers, instruments, databases, conversations, institutions, and machines.

This has direct implications for artificial intelligence. Current AI systems already demonstrate impressive forms of pattern completion, language generation, summarization, coding, analogy, and problem solving. They can answer questions by drawing on enormous statistical structure encoded during training and made available through context. However, answering a prompt is not the same thing as maintaining an autonomous line of thought. A system can generate plausible responses without preserving an unresolved problem as a self-updating internal state.

A more creative artificial system would require persistent query construction. It would maintain unresolved problems over time. It would form composite working-memory states containing goals, constraints, evidence, partial hypotheses, analogies, failures, and evaluation criteria. It would use these states to generate updates. It would incorporate the best updates into revised query states. It would continue this process across many iterations, not merely within a single response, but across extended periods of autonomous investigation.

Such a system would not only produce answers. It would produce better questions. It would not only search a database or latent space. It would construct new search states. It would not merely recombine information randomly. It would recombine information under constraints supplied by goals, evidence, coherence, predictive success, and utility. It would not merely generate novelty. It would evaluate, stabilize, and build on novelty.

This distinction may be central to superintelligence. A superintelligent system would not be defined only by more memory, faster computation, larger models, or broader training data. Those capacities would matter, but they would not be sufficient. A system capable of sustained discovery would need to carry forward unresolved problems, transform them through intermediate products, and recursively deepen the query space. Superintelligence may therefore require not only superior answers, but superior iterative query construction.

This does not mean that human thought and artificial thought must use identical mechanisms. Biological working memory depends on neural activity, synaptic potentiation, embodiment, affect, development, and evolution. Artificial systems may implement analogous functions using different substrates. The relevant abstraction is not the biological mechanism alone, but the computational role: a temporary state must preserve selected prior contents, incorporate useful updates, and use the revised state to guide subsequent search.

If artificial systems are built to perform this process at scales beyond human capacity, they may generate forms of discovery that exceed ordinary human innovation. They could maintain thousands of unresolved problems, explore many query trajectories in parallel, preserve intermediate products indefinitely, test them against vast bodies of data, and recombine them across domains that no human expert could simultaneously master. Such systems would not merely automate research tasks. They would automate the construction of research questions.

This is why the phrase iterative query construction may be useful beyond cognitive theory. It identifies a capability that any discovery-oriented intelligence may need. A system that cannot revise its own query state remains dependent on externally supplied prompts, goals, or problem frames. A system that can revise its own query state can participate in the creation of those frames. It can move from answering to investigating, from investigating to theorizing, and from theorizing to discovery.

The central implication is that creativity and innovation do not require a separate faculty outside ordinary cognition. They require the extension, stabilization, and refinement of the same process that generates thought moment by moment. A mind becomes innovative when it can preserve unresolved questions, construct unfamiliar but constrained query states, incorporate intermediate products, and continue until a useful new representation emerges. A culture becomes innovative when it can preserve and recombine these products across minds. A superintelligent system may become scientifically transformative when it can perform this process autonomously, recursively, and at scale.

8. Predictions, Tests, and Conclusion

The hypothesis that thinking is continuous iterative query construction is intended to be more than a metaphor. It should generate empirical, computational, and phenomenological predictions. If working-memory states function as self-modifying composite queries, then the structure of thought should exhibit measurable patterns of overlap, transformation, constraint, and update selection.

The first prediction is that consecutive cognitive states should show representational overlap. The model predicts that a line of thought should not consist of discrete, unrelated states. Instead, successive states should preserve a subset of active or recently active contents. In neural terms, this may appear as overlapping population codes, persistent activity, short-term synaptic potentiation, recurrent activation patterns, or sustained representational similarity across adjacent time windows. In psychological terms, it should appear as continuity of topic, goal, scene, premise, or problem frame.

The second prediction is that the amount of overlap should vary with cognitive function. Simple recall may require less extended overlap than multi-step reasoning. Complex planning should require longer persistence of goal and constraint representations. Creative cognition should require enough overlap to preserve the problem, but enough replacement to introduce remote or reweighted associations. The model therefore predicts an optimal intermediate regime. Too little retention should produce fragmentation. Too much retention should produce fixation.

The third prediction is that creative thought should involve longer chains of partially overlapping intermediate states than simple retrieval. In a recall task, a familiar cue may quickly converge on a stored answer. In a creative task, an initial query state should produce intermediate products that then become part of subsequent query states. This should be detectable behaviorally as sequences of partial solutions, revisions, analogies, false starts, reformulations, and reweightings of relevance. It may also be detectable neurally as evolving but partially preserved representational trajectories.

The fourth prediction is that insight should occur when a new representation changes the relevance structure of retained contents. An insight is not merely the arrival of a new item. It is an update that reorganizes the query. Before insight, the active coalition may fail to converge on a useful continuation. After the critical update, previously retained contents become newly related. The same problem elements now point toward a different solution. This predicts that moments of insight should involve abrupt changes in the weighting of prior representations rather than the simple addition of an unrelated answer.

The fifth prediction is that fixation and perseveration should reflect excessive retention or insufficient query reconstruction. When a thinker remains trapped in the same interpretation, strategy, or association, the active coalition may keep selecting the same class of updates. In such cases, the query is stable but unproductive. Breaking fixation should require altering the query state, either by suppressing dominant associations, introducing a new representation, changing the goal, shifting modality, or externalizing the problem in a new form.

The sixth prediction is that distraction should reflect uncontrolled query capture. A distracting update enters working memory and changes the query away from the prior goal. The issue is not that the system updates, but that the update fails to preserve the constraints necessary for the original trajectory. Productive thought requires updates that transform the query while maintaining enough continuity with the task. Distraction is transformation without the right continuity.

The seventh prediction is that expertise should alter the structure of query construction. Experts should be able to form more compressed, abstract, and high-value query states because their long-term memory contains richer representations and more useful constraints. A single expert-level concept may stand for a large body of relations that a novice would need many separate items to represent. This should allow experts to search more efficiently and to generate more powerful recombinations. However, expertise may also increase fixation when dominant schemas overconstrain the query.

The eighth prediction concerns artificial systems. An AI architecture that maintains a persistent, capacity-limited, partially updated working-memory state should outperform a fixed-query system on tasks requiring extended reasoning, hypothesis generation, conceptual synthesis, open-ended problem reformulation, and creative discovery. Such a system should not merely generate one output from one input. It should preserve selected intermediate products, incorporate them into future internal states, and continue searching from the revised state. Its performance should depend on the balance between retention and replacement.

These predictions suggest several possible tests. Neuroimaging studies could examine representational overlap across consecutive moments of reasoning or creative problem solving. Electrophysiological studies could test whether insight involves reweighting of retained representations. Behavioral studies could compare tasks that require simple recall, sequential reasoning, divergent thinking, and innovation, measuring the number and structure of intermediate states. Computational models could compare fixed-query retrieval, ordinary chain-of-thought generation, and architectures with explicit partially updated working-memory states.

One important methodological challenge is that the relevant states may not correspond neatly to verbal reports. Many query states are likely to be subverbal, imagistic, affective, motoric, or distributed across multiple modalities. A participant’s report of a thought may capture only the most salient or reportable component of a richer composite query. For this reason, tests of the model should not rely only on introspection or language. They should combine behavioral traces, verbal protocols, neural measurements, and computational modeling.

A second challenge is that the boundaries between states may be artificial. The model uses the language of states and updates, but the underlying process may be continuous. Any attempt to measure it must impose time windows, sampling intervals, or representational snapshots. This does not undermine the model. It reflects the difficulty of studying a continuous process using discrete measurements. The central prediction is not that thought proceeds in perfectly separated steps, but that adjacent windows of thought should show structured partial overlap and progressive transformation.

A third challenge is that multiple mechanisms may implement query construction. Sustained firing, synaptic potentiation, oscillatory coordination, recurrent loops, attentional gain, priming, and external symbolic artifacts may all contribute. The model does not require that only one mechanism be responsible. It requires that cognitive systems preserve selected prior contents, introduce new contents, and use the revised coalition to constrain subsequent updates. Different organisms, brain regions, developmental stages, tasks, and artificial systems may implement this functional pattern differently.

The theoretical value of continuous iterative query construction lies in its ability to connect several levels of explanation. At the neural level, it relates thought to persistent activity, potentiation, and recurrent dynamics. At the cognitive level, it relates thought to working memory, attention, association, prediction, and control. At the phenomenological level, it explains why thought feels continuous yet changeable. At the creativity level, it explains how original ideas can emerge from familiar materials. At the AI level, it suggests a design principle for systems capable of autonomous discovery.

The account also helps clarify why thought is difficult to reduce to a sequence of outputs. A spoken sentence, written line, motor action, or explicit answer is only the visible edge of a deeper process. Beneath the output is a continuously revised search state. The mind is not merely producing responses. It is maintaining, transforming, and querying an evolving internal context. Outputs matter because they can stabilize parts of that context, communicate it to others, and allow it to reenter thought as an externalized query element.

The proposed model also reframes the relation between continuity and originality. Continuity is often associated with preservation, while originality is associated with change. But in this account, originality depends on continuity. Intermediate products must persist long enough to combine with later products. A problem frame must remain active long enough for a remote association to become relevant. A new analogy must be held long enough to reorganize the target domain. Without retention, there is no compounding. Without replacement, there is no transformation. Thought requires both.

The conclusion is therefore straightforward. Thinking is continuous iterative query construction. The active contents of working memory form a composite query. Multiassociative search selects an update. Iterative updating incorporates that update while retaining part of the prior state. The revised state becomes the next query. Through repeated cycles of retention, replacement, and recombination, the mind recalls, reasons, imagines, plans, creates, and discovers.

This account does not require dismissing existing cognitive models. Many of them describe genuine components of cognition. The present proposal attempts to specify the transition rule that allows those components to operate as a moving stream. A mind is not only a system that stores information, broadcasts contents, predicts inputs, associates ideas, or searches problem spaces. A mind is a system that continuously reconstructs the query from which its next state will be selected.

If this hypothesis is correct, then creativity is not an inexplicable exception to ordinary cognition. It is what happens when the ordinary process of thought is extended, stabilized, and constrained across enough iterations to produce a new and useful representation. Innovation is what happens when that representation survives testing and enters shared memory. Superintelligence may arise when artificial systems can perform this process autonomously, recursively, and at scales unavailable to human minds.

The deepest function of thought may not be to find answers to fixed questions. It may be to build better questions while searching for answers. In that sense, intelligence is not merely the capacity to search. It is the capacity to transform the search state itself.

References

Anderson, J. R. (1983). A spreading activation theory of memory. Journal of Verbal Learning and Verbal Behavior, 22(3), 261-295.

Anderson, J. R. (1996). ACT: A simple theory of complex cognition. American Psychologist, 51(4), 355-365.

Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: A proposed system and its control processes. In K. W. Spence & J. T. Spence (Eds.), The psychology of learning and motivation (Vol. 2, pp. 89-195). Academic Press.

Baars, B. J. (1988). A cognitive theory of consciousness. Cambridge University Press.

Baars, B. J. (2005). Global workspace theory of consciousness: Toward a cognitive neuroscience of human experience. Progress in Brain Research, 150, 45-53.

Baars, B. J., & Franklin, S. (2007). An architectural model of conscious and unconscious brain functions: Global Workspace Theory and IDA. Neural Networks, 20(9), 955-961.

Baddeley, A. D. (1986). Working memory. Clarendon Press.

Baddeley, A. D. (2000). The episodic buffer: A new component of working memory? Trends in Cognitive Sciences, 4(11), 417-423.

Baddeley, A. D. (2007). Working memory, thought, and action. Oxford University Press.

Baddeley, A. D. (2012). Working memory: Theories, models, and controversies. Annual Review of Psychology, 63, 1-29.

Baddeley, A. D., & Hitch, G. J. (1974). Working memory. In G. A. Bower (Ed.), Recent advances in learning and motivation (Vol. 8, pp. 47-89). Academic Press.

Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford University Press.

Botvinick, M. M. (2008). Hierarchical models of behavior and prefrontal function. Trends in Cognitive Sciences, 12(5), 201-208.

Botvinick, M. M., & Plaut, D. C. (2006). Short-term memory for serial order: A recurrent neural network model. Psychological Review, 113(2), 201-233.

Braver, T. S., & Cohen, J. D. (2000). On the control of control: The role of dopamine in regulating prefrontal function and working memory. In S. Monsell & J. Driver (Eds.), Attention and performance XVIII: Control of cognitive processes (pp. 713-737). MIT Press.

Christophel, T. B., Klink, P. C., Spitzer, B., Roelfsema, P. R., & Haynes, J. D. (2017). The distributed nature of working memory. Trends in Cognitive Sciences, 21(2), 111-124.

Collins, A. M., & Loftus, E. F. (1975). A spreading-activation theory of semantic processing. Psychological Review, 82(6), 407-428.

Constantinidis, C., Funahashi, S., Lee, D., Murray, J. D., Qi, X. L., Wang, M., & Arnsten, A. F. T. (2018). Persistent spiking activity underlies working memory. Journal of Neuroscience, 38(32), 7020-7028.

Cowan, N. (1988). Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information-processing system. Psychological Bulletin, 104(2), 163-191.

Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87-114.

Cowan, N. (2005). Working memory capacity. Psychology Press.

Cowan, N. (2009). What are the differences between long-term, short-term, and working memory? Progress in Brain Research, 169, 323-338.

D’Esposito, M., & Postle, B. R. (2015). The cognitive neuroscience of working memory. Annual Review of Psychology, 66, 115-142.

Damasio, A. R. (1989). Time-locked multiregional retroactivation: A systems-level proposal for the neural substrates of recall and recognition. Cognition, 33(1-2), 25-62.

Dehaene, S. (2014). Consciousness and the brain: Deciphering how the brain codes our thoughts. Viking.

Eliasmith, C. (2013). How to build a brain: A neural architecture for biological cognition. Oxford University Press.

Fauconnier, G., & Turner, M. (2002). The way we think: Conceptual blending and the mind’s hidden complexities. Basic Books.

Finke, R. A., Ward, T. B., & Smith, S. M. (1992). Creative cognition: Theory, research, and applications. MIT Press.

Fuster, J. M. (1973). Unit activity in prefrontal cortex during delayed-response performance: Neuronal correlates of transient memory. Journal of Neurophysiology, 36(1), 61-78.

Fuster, J. M. (2005). Cortex and mind: Unifying cognition. Oxford University Press.

Fuster, J. M. (2009). Cortex and memory: Emergence of a new paradigm. Journal of Cognitive Neuroscience, 21(11), 2047-2072.

Fuster, J. M. (2015). The prefrontal cortex (5th ed.). Academic Press.

Goldman-Rakic, P. S. (1995). Cellular basis of working memory. Neuron, 14(3), 477-485.

Goertzel, B., Pennachin, C., & Geisweiller, N. (2014). Engineering general intelligence: A path to advanced AGI via embodied learning and cognitive synergy. Atlantis Press.

Hebb, D. O. (1949). The organization of behavior: A neuropsychological theory. Wiley.

Hofstadter, D. R. (2007). I am a strange loop. Basic Books.

James, W. (1890). The principles of psychology. Henry Holt.

James, W. (1909). A pluralistic universe. Longmans, Green, and Co.

Laird, J. E. (2012). The Soar cognitive architecture. MIT Press.

Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167-202.

Minsky, M. (1986). The society of mind. Simon & Schuster.

Newell, A., & Simon, H. A. (1972). Human problem solving. Prentice-Hall.

Oberauer, K. (2002). Access to information in working memory: Exploring the focus of attention. Journal of Experimental Psychology: Learning, Memory, and Cognition, 28(3), 411-421.

Postle, B. R. (2006). Working memory as an emergent property of the mind and brain. Neuroscience, 139(1), 23-38.

Reser, J. E. (2016). Incremental change in the set of coactive cortical assemblies enables mental continuity. Physiology & Behavior, 167, 222-237.

Reser, J. E. (2022-2024). A cognitive architecture for machine consciousness and artificial superintelligence: Updating working memory iteratively. arXiv:2203.17255 [q-bio.NC].

Rose, N. S., LaRocque, J. J., Riggall, A. C., Gosseries, O., Starrett, M. J., Meyering, E. E., & Postle, B. R. (2016). Reactivation of latent working memories with transcranial magnetic stimulation. Science, 354(6316), 1136-1139.

Stokes, M. G. (2015). Activity-silent working memory in prefrontal cortex: A dynamic coding framework. Trends in Cognitive Sciences, 19(7), 394-405.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.

Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T. L., Cao, Y., & Narasimhan, K. (2023). Tree of thoughts: Deliberate problem solving with large language models. arXiv:2305.10601.