A Working-Memory Model of Originality, Hypothesis Generation, and Autonomous Scientific Thought

Abstract

Creativity is often defined as the production of something novel and useful. This definition describes a property of creative products, but it does not explain the cognitive process by which useful novelty is generated. This article proposes that creativity is recursive search-space construction. On this view, creative cognition does not merely search within a predefined space of possible answers. It progressively constructs and reconstructs the space in which new answers, hypotheses, tasks, and discoveries become reachable.

The model builds on the concepts of state-spanning coactivity, incremental change in state-spanning coactivity, polyassociativity, and iterative updating of working memory. Working memory is treated as the active site of creative construction: a temporary coalition of perceptual, semantic, affective, motoric, mnemonic, and goal-related representations. This coalition defines the current problem state. Through multiassociative search, the active coalition converges on candidate updates. Through incremental updating, some prior contents are retained, others are removed or inhibited, and new contents are incorporated. The revised state then defines a new search space.

This framework distinguishes novelty from creativity. Novelty can be generated by randomness, hallucination, error, or arbitrary recombination. Creativity requires novelty under constraint. Creative updates are valuable when they reorganize the active problem state, preserve relevant constraints, introduce useful variation, and generate new questions, hypotheses, tests, or tasks. Insight is interpreted as a state change in which a new representation reorganizes the relevance structure of a previously unproductive search space. Incubation is interpreted as altered search-space readiness, in which unresolved problem elements remain latent, potentiated, or more easily reactivated.

The model also reframes scientific discovery and artificial intelligence. A hypothesis is not merely a possible answer. It is a search-space modifier that changes which evidence matters, which tests should be performed, and which new problems can be formulated. Current artificial intelligence systems are increasingly capable of output generation, tool use, long-horizon task execution, and partial scientific workflow automation. However, autonomous scientific creativity requires more than completing assigned tasks. It requires persistent unresolved-problem memory, anomaly detection, curiosity-like gating, working-memory-like problem states, multimodal rendering and testing, recursive incorporation of intermediate products, and task invention.

The central thesis is that creativity is not a separate magical faculty. It is a special use of ordinary cognitive mechanisms operating under conditions of novelty, constraint, persistence, and evaluation. A mind creates by preserving enough of the old problem to remain coherent, introducing enough new material to transform it, and recursively building the space in which an original answer, hypothesis, task, or discovery can become possible.

Keywords

creativity; recursive search-space construction; working memory; state-spanning coactivity; incremental change in state-spanning coactivity; polyassociativity; multiassociative search; iterative updating; originality; insight; incubation; hypothesis generation; scientific discovery; task invention; autonomous scientific thought; artificial intelligence; AI agents; AI scientist; machine creativity; superintelligence.

1. Introduction: Creativity Beyond Output Generation

Creativity is often defined as the production of something novel and useful. This definition is serviceable, but incomplete. It describes a property of the product rather than the process that generates the product. It tells us that a creative idea must be new and valuable, but it does not explain how a cognitive system constructs the conditions under which such an idea can arise.

This distinction matters because novelty by itself is easy to generate. A random process can produce novelty. A dream can produce novelty. A hallucination can produce novelty. A language model can produce unusual combinations of words. A person can make arbitrary associations. But creativity is not mere novelty. Creative cognition produces novelty that is constrained, interpretable, useful, explanatory, beautiful, predictive, testable, or action-guiding. It generates something new that fits some problem, goal, medium, reality, or evaluative context.

The central claim of this article is that creativity is recursive search-space construction. A creative mind does not merely search inside a fixed space of possible answers. It progressively constructs and reconstructs the search space itself. It begins with a temporary cognitive state composed of memories, goals, analogies, constraints, percepts, emotions, unresolved tensions, and partial hypotheses. This state determines what can be searched, what counts as relevant, what kinds of updates are likely to occur, and what would count as progress. When a new idea, image, association, hypothesis, or question enters working memory, it changes the state from which the next search will proceed. Creativity emerges when this cycle recursively transforms the problem space.

This framing builds on the view that thought proceeds through the incremental updating of working memory. Consecutive mental states are not fully independent. Some representations persist across state transitions, while others fade, are inhibited, or are replaced. Newly activated representations enter into coactivity with representations that remain from the previous state. Because of this partial overlap, each state is continuous with the state before it, but not identical to it. Thought can therefore preserve a problem while transforming it.

In ordinary recall, a familiar cue may retrieve a familiar completion. In ordinary problem solving, a predefined task may guide a sequence of operations toward a known type of answer. In creative thought, however, the task itself may be incomplete, unstable, or unnamed. The thinker may not yet know what question should be asked. The initial state may contain only a tension, anomaly, intuition, image, constraint, or sense that something important has not yet been articulated. The creative process then becomes the construction of a search space in which a more precise question, hypothesis, or task can emerge.

This is especially important for scientific creativity. A scientific discovery rarely begins as a clean question with a known form of answer. It may begin as an anomaly that does not fit an existing theory, an unexplained pattern, an analogy between distant domains, a methodological possibility, or a vague dissatisfaction with a prevailing interpretation. The creative act involves forming a problem representation that makes the anomaly searchable. A hypothesis is not merely an answer. It is a transformation of the search space. Once a hypothesis is introduced, it changes which observations matter, which tests should be run, which alternatives should be considered, and which subproblems should be named.

The same logic applies to artistic, technological, mathematical, and philosophical creativity. A painter does not merely retrieve a finished image. A composer does not merely select a melody from memory. An engineer does not merely locate an invention in a preexisting catalog of possibilities. A mathematician does not simply search a fixed space of proofs. In each case, intermediate products alter the space of possible next moves. A mark on the canvas changes the composition. A motif changes the harmonic expectation. A prototype reveals a hidden constraint. A lemma changes the proof space. A distinction changes the conceptual field. Creative work proceeds because each partial product reshapes what can be searched next.

This also reframes the challenge of artificial intelligence. Many contemporary AI systems are increasingly capable of producing impressive outputs and executing complex tasks when prompted. They can summarize, code, plan, search, generate images, solve problems, and coordinate tool use. But output generation and task completion are not the same as autonomous creativity. A system that answers supplied questions is not yet necessarily a system that can decide which questions need to be asked. A system that completes assigned tasks is not yet necessarily a system that can originate the tasks whose completion would matter.

A genuinely creative artificial system would need to maintain unresolved problems, construct internal problem states, generate hypotheses, identify missing evidence, propose tests, interpret failures, and revise its own search space. It would need to turn intermediate products into new cognitive conditions. It would not merely respond to a prompt. It would construct the next prompt for itself.

The present article proposes a working-memory model of this capacity. Creative thought depends on temporary coalitions of active and recently potentiated representations. These coalitions define the current search space. Through multiassociative search, the coalition converges on candidate updates. Through incremental updating, some prior contents are retained, some are removed, and new contents are incorporated. The resulting state defines a revised search space. Through repetition, this process can produce originality, insight, hypothesis generation, task invention, and autonomous scientific thought.

The central thesis can be stated simply: creativity is not only the generation of novel outputs. It is the recursive construction of the search spaces from which novel and useful outputs become possible.

2. Novelty Is Not Enough

Any adequate theory of creativity must distinguish novelty from creative value. Novelty can be produced by randomization, error, noise, mutation, hallucination, or arbitrary recombination. A string of unrelated words is novel. A nonsensical drawing may be novel. A mistaken inference may be novel. A false hypothesis may be novel. But novelty alone does not constitute creativity. A creative product must be new in a way that is constrained by a larger structure of relevance.

Those constraints can take many forms. In science, a creative idea must be constrained by evidence, prediction, mechanistic plausibility, explanatory scope, and testability. In engineering, it must be constrained by function, materials, cost, energy, reliability, and implementation. In mathematics, it must be constrained by definitions, axioms, transformations, and proof. In art, it may be constrained by medium, form, emotional resonance, composition, style, cultural meaning, or aesthetic coherence. In ordinary life, it may be constrained by social context, personal goals, practical limitations, and the anticipated consequences of action.

Creativity therefore requires a balance between variation and constraint. Variation allows the system to escape familiar completions. Constraint prevents the system from degenerating into incoherence. Too little variation produces repetition, fixation, and conventionality. Too little constraint produces randomness, irrelevance, and hallucination. Creative thought occupies the intermediate regime in which the system can depart from the obvious while remaining answerable to something.

This balance is difficult to explain if creativity is treated as simple association. If one idea merely calls up another, then creativity becomes a matter of reaching remote associations. Remote association is important, but it is insufficient. A remote association may be useless. It may be funny, distracting, poetic, or false. It becomes creative only if it reorganizes the problem in a productive way. The association must do work. It must change the search space so that new possibilities become available.

The same problem appears if creativity is treated as divergent output generation. A system can generate many candidate outputs without being creative in the stronger sense. Quantity of variation does not guarantee quality of transformation. A creative system must not only produce candidates. It must maintain the problem, evaluate candidate updates, preserve useful intermediate products, discard misleading ones, and use the surviving products to change what is searched next. Creativity is not merely divergence. It is recursive reconstruction under constraint.

This is where working memory becomes central. The constraints that make novelty useful must be maintained somewhere. A creative system must keep the problem frame active enough that new material can be judged against it. At the same time, the frame must remain flexible enough to be revised. If the system cannot preserve the constraints, thought becomes fragmented. If it cannot revise the constraints, thought becomes fixated. Creativity depends on the controlled transformation of an active problem state.

Consider the difference between a hallucination and an insight. Both may involve an unexpected association. But in hallucination, the association may not remain accountable to the surrounding constraints. It may produce novelty without integration. In insight, the unexpected representation reorganizes the current problem state. It makes previously disconnected elements fit together. It reveals why some evidence mattered, why a prior strategy failed, or why a new test should be performed. The insight is not only a new content. It is a new organization of relevance.

This distinction is especially important for artificial intelligence. Large generative models can produce outputs that appear novel, fluent, and surprising. But a generated output is not creative simply because it is unusual. The critical question is whether the system can evaluate the output against a persistent problem state and then use the result to revise its own future inquiry. Without that recursive process, novelty remains shallow. It may be impressive as production, but it does not yet amount to autonomous creative thought.

A creative AI system would need more than a generator. It would need a mechanism for retaining unresolved constraints, detecting when an output changes the problem, incorporating useful changes into memory, and creating new tasks from the implications of those changes. It would need to ask: What does this imply? What does it make newly relevant? What should be tested next? What problem has this answer revealed? What question does this partial result allow me to formulate now that I could not formulate before?

Human creativity appears to operate in this way. A person working on a theory, painting, invention, or proof does not merely produce independent variants. Each partial product alters the next attempt. A failed formulation clarifies a constraint. An analogy makes a new relation salient. A sketch reveals an imbalance. A prototype exposes an implementation problem. A preliminary hypothesis suggests a discriminating test. Creativity advances because intermediate products are not discarded as isolated outputs. They are folded back into the active search space.

This recursive incorporation is what makes novelty productive. A creative system must be able to generate unexpected material, but it must also preserve the parts of that material that change the structure of the problem. It must be able to relax constraints temporarily, decompose familiar chunks, and allow unusual combinations to form. But it must also be able to reimpose constraint, evaluate coherence, and stabilize what remains useful. The creative process alternates between loosening and tightening, expansion and selection, exploration and consolidation.

The working-memory model developed here provides a mechanism for this alternation. Active representations define the current search space. Novel candidate updates enter the active state. Some updates are rejected because they do not fit the retained constraints. Others are incorporated because they make the search space more informative. Once incorporated, they change the future trajectory of thought. Creativity is therefore not novelty alone, but novelty that survives incorporation into a recursively developing problem state.

The central claim of this section is that creativity cannot be reduced to new output generation. New outputs become creative only when they participate in a constrained sequence of search-space transformation. The mind creates by maintaining enough continuity to preserve the problem, introducing enough variation to transform it, and recursively building a space in which an original result can be found.

3. Working Memory as the Site of Creative Construction

Creativity requires a temporary site in which representations can be brought together, held active, modified, evaluated, and recombined. That site is working memory. Working memory is not merely a passive store for information that has already been selected. It is the active arena in which the current problem state is assembled. The contents of working memory define what is presently being considered, what is being ignored, what remains unresolved, and what kinds of updates would count as relevant.

In creative thought, working memory does not hold a single idea. It holds a coalition. This coalition may include a goal, an image, an analogy, a prior failure, an emotional valuation, a remembered fact, a perceptual cue, a methodological constraint, a partial hypothesis, and a vague sense that something does not yet fit. The coalition is heterogeneous. It may contain verbal, visual, motoric, affective, semantic, episodic, and procedural elements at the same time. These elements are not simply juxtaposed. They interact. Together, they determine the active search space.

This is why creativity cannot be explained only as retrieval from long-term memory. Long-term memory supplies the stored materials, but working memory determines which materials are coactive at a particular moment. The same memory can participate in many different creative states depending on the other contents with which it is combined. A concept that is ordinary in one context can become generative in another because the surrounding coalition changes its implications.

For example, the concept of “selection” means something different when it is coactive with evolutionary biology, machine learning, attention, market behavior, immune response, or artistic editing. The concept itself is not creative by itself. It becomes creatively useful when it enters a coalition that allows relations across domains to become active together. Working memory supplies the temporary frame in which this can happen.

The creative search space is therefore not simply the total space of everything known by the system. It is the locally active subset of knowledge, imagery, goal structure, and constraint that is presently able to interact. A system may possess vast stored knowledge and still fail to create if it cannot place the right representations into coactivity. Conversely, a small number of well-chosen active representations can generate a powerful search space if their relations are sufficiently rich.

Working memory also allows creativity because it can make nonsimultaneous things simultaneous. Many important relations are distributed through time. A cause may precede an effect by seconds, years, or generations. An observation made in one domain may become relevant to a theory in another domain much later. A failed attempt may only become meaningful after a new analogy appears. By sustaining representations after their original inputs have passed, working memory allows temporally separated elements to become coactive inside the mind. This ability is central to creative cognition.

A creative idea often begins when two or more representations that did not occur together in the environment are made to occur together internally. The brain can hold a prior observation active while a new concept enters awareness. It can hold a goal active while several candidate solutions are tried. It can hold a contradiction active while searching for a hidden assumption. It can preserve the memory of a failed explanation while a new mechanism is introduced. In each case, working memory creates an artificial simultaneity that the external world did not provide.

This artificial simultaneity is not arbitrary. The system does not benefit from placing every representation into coactivity with every other representation. The active coalition must be selective. It must preserve what matters and exclude what interferes. Creative construction depends on the capacity to maintain a problem frame while allowing its contents to change. Too narrow a coalition produces fixation. Too broad a coalition produces confusion. A productive creative state is constrained enough to remain meaningful, but open enough to admit transformation.

Working memory also gives creative thought its continuity. A creative project cannot be completed if every moment is disconnected from the previous one. A hypothesis must remain active long enough to generate predictions. A visual form must remain stable long enough to be modified. A musical theme must persist long enough to be varied. A philosophical distinction must remain available long enough to reorganize the concepts around it. Creativity requires that intermediate products survive long enough to be used.

At the same time, working memory must not preserve everything. If the active state is too stable, the same associations keep recurring. The system repeats familiar completions. A problem is interpreted through the same frame again and again. In this case, working memory becomes a mechanism of fixation rather than creativity. Creative construction requires selective retention and selective replacement. Some representations must be conserved because they define the problem. Others must be released because they prevent the problem from being reformulated.

This provides a way to understand the role of constraint relaxation in creativity. Creative thought often requires loosening the active problem state. A familiar chunk may need to be decomposed. A habitual interpretation may need to be inhibited. A dominant association may need to lose priority. A peripheral feature may need to become central. Constraint relaxation does not mean abandoning structure altogether. It means altering the structure of the active search space so that a different class of updates can become available.

Working memory is also multimodal. Creative thought is not restricted to verbal propositions. The mind can construct visual scenes, motor simulations, auditory patterns, affective tones, spatial layouts, social scripts, and abstract relations. These modalities can interact. A verbal hypothesis can generate an image. An image can reveal a structural relation. A motor simulation can expose an implementation problem. A feeling of mismatch can guide attention toward a hidden inconsistency. A creative working-memory state is often a multimodal workspace in which different representational systems constrain one another.

This multimodality matters because creative products often emerge from rendering. A high-level idea may be underspecified until it is turned into a sentence, sketch, diagram, equation, simulation, experiment, movement, or prototype. Rendering forces the idea to interact with the constraints of a medium. The medium then returns information. A sentence reveals ambiguity. A diagram reveals symmetry. A prototype reveals friction. A simulation reveals instability. A sketch reveals proportion. The rendered product becomes a new input to working memory and changes the search space.

In this sense, working memory is not a solitary inner screen. It is an active interface between internal representation and externalization. A thinker can use paper, language, tools, diagrams, code, instruments, and other people to stabilize intermediate products. Once stabilized, those products can reenter working memory as new constraints. This extends creative cognition beyond the biological brain, but it does not replace working memory. External artifacts matter because they help working memory preserve, inspect, modify, and recombine representations that would otherwise decay.

The same principle applies to artificial systems. A creative AI would not merely need a large memory store or a powerful generator. It would need a persistent active state in which unresolved problems, constraints, candidate hypotheses, observations, failures, and evaluative criteria can be co-maintained and revised. It would need to render partial products into language, code, diagrams, simulations, experiments, or plans, then feed the results back into the active state. Without such a working-memory-like process, outputs may remain isolated products rather than steps in a developing creative trajectory.

Thus, working memory is the site of creative construction because it defines the active search space. It makes selected representations simultaneously available. It preserves continuity across time. It allows constraints to be relaxed and reimposed. It supports multimodal rendering and feedback. It lets intermediate products become resources for subsequent thought. Creativity occurs when working memory does not merely hold an idea, but recursively constructs and reconstructs the space in which new ideas become reachable.

4. From State-Spanning Coactivity to Recursive Search-Space Construction

The claim that creativity is recursive search-space construction can be grounded in the earlier concepts of state-spanning coactivity and incremental change in state-spanning coactivity. State-spanning coactivity refers to the fact that some neural representations remain coactive across successive brain states. Incremental change in state-spanning coactivity refers to the gradual turnover of this active set, where some contents persist, some are deactivated, and some newly active contents enter. This is the biological foundation for a cognitive process that is both continuous and transformative.

If all contents of working memory were replaced at once, thought would lose continuity. Each state would be isolated from the one before it. If no contents were replaced, thought would become static. The system would preserve a state but fail to progress. Creativity requires the intermediate pattern: partial persistence and partial replacement. Retained representations preserve the problem. Newly introduced representations transform it. The creative process depends on the fact that the system can remain the same enough to maintain a search space, while changing enough to reconstruct that space.

The retained contents of state-spanning coactivity can be understood as the active core of the problem. They may include the goal, topic, anomaly, question, image, hypothesis, or constraint that gives the sequence its identity. New contents enter into relation with this retained core. They do not appear in an empty field. They are interpreted against what remains active from the preceding state. This is why a new representation can have creative force. It arrives into a structured context and changes the structure.

This mechanism gives creativity a recursive character. A state of working memory generates or recruits an update. The update is incorporated into the remaining contents of the prior state. The resulting coalition becomes the next active problem state. That state then recruits another update. The output of one cycle becomes part of the input to the next. The search space is therefore not fixed in advance. It is constructed recursively by the sequence of updates it helps produce.

This differs from ordinary search. In ordinary search, the space of possible answers is often assumed to be defined before the search begins. The system searches within that space until it finds a solution. In creative cognition, the space itself changes as the process unfolds. A new analogy changes what counts as similar. A new hypothesis changes what evidence matters. A new image changes what form seems possible. A new distinction changes the conceptual landscape. A new failure reveals a hidden constraint. The search space is constructed by the very act of searching.

The earlier concept of polyassociativity provides the selection mechanism. Polyassociativity means that several coactive representations jointly converge on the next representation to enter the active state. One idea does not simply call up the next in a linear chain. Rather, a coalition of active representations pools its influence across the network and activates a representation that is most relevant to the group as a whole. At the cognitive level, this means that a goal, memory, analogy, image, and constraint can jointly determine the next thought.

This is essential for creativity. A creative update is usually not the strongest associate of a single idea. It is often the result of convergent pressure from several partially related contents. One representation points toward a domain. Another supplies a constraint. Another supplies an analogy. Another supplies a missing mechanism. Another supplies an evaluative standard. The update that enters working memory may be weakly associated with each item individually, but highly relevant to the coalition as a whole.

A novel convergence event occurs when a group of representations that has not previously been active together converges on a useful update. This is one of the most important mechanisms for creative thought. The component representations may all be familiar. The update may even be familiar in other contexts. But the coalition is new, and the role played by the update may be new. Creativity begins when familiar representations are placed into a novel configuration that causes the system to converge on a representation, relation, or hypothesis that was not previously accessible from any one of them alone.

Novel convergence events can explain why creative ideas often feel both surprising and appropriate. They are surprising because the active coalition is unusual. They are appropriate because the update is not random. It is selected by convergent support from the coalition. This gives creativity its characteristic combination of novelty and fit. The idea was not obvious, but once it appears, it seems connected to the problem.

However, a single novel convergence event is not enough to explain extended creativity. The critical step is incorporation. Once the new representation enters working memory, it changes the next search space. A creative update becomes powerful when it is not merely noticed, but retained and used. It may become a new premise, analogy, image, constraint, or hypothesis. It may cause older elements to be reweighted. It may reveal a hidden relation among retained contents. It may imply a new test or subproblem. In this way, the novel convergence event becomes a recursive search-space modifier.

This allows creative products to compound. One update introduces a candidate relation. The next update tests or elaborates it. A third update reframes the original problem. A fourth update names a new subproblem. A fifth update suggests an experiment. The final creative product may appear unified, but it is assembled from a chain of partially overlapping states. Each state carries forward selected contents from earlier states and adds something new.

This model also explains why creativity often requires sustained attention. A novel convergence event may be fragile. If the new representation is not stabilized, it may disappear before its implications can be explored. The system must hold it in coactivity with the original problem long enough for further updates to occur. This is why insight often needs elaboration. The moment of insight may introduce the crucial representation, but creative work continues as the system discovers what the insight changes.

At the same time, creativity requires the ability to disrupt the current state. State-spanning coactivity preserves continuity, but excessive continuity can produce fixation. If the same representations remain active for too long, the same search space persists and the same class of updates recurs. Creative thought may require a partial reduction of continuity so that dominant chunks can decompose, familiar associations can weaken, and new representations can enter. The optimal creative state may therefore involve a dynamic balance between persistence and turnover.

The balance between persistence and turnover can vary across phases of creativity. During problem formulation, the system may need broad exploration and flexible representation. During hypothesis elaboration, it may need sustained focus. During evaluation, it may need stricter constraint. During incubation, focal contents may fade while latent elements remain potentiated. During insight, a new representation may reorganize a partially preserved problem state. During implementation, the system may return to more controlled and stable search.

This framework also helps explain why creative thought is often assisted by changing modality. A verbal problem may become more tractable when drawn. A visual design may become clearer when described. A theory may become sharper when formalized mathematically. A proposed mechanism may become more constrained when simulated. Each modality renders the problem differently. Each rendering creates new active features. Those features reenter working memory and alter the search space. Modality shifts are therefore not incidental. They are mechanisms of recursive search-space construction.

For artificial intelligence, this suggests a concrete design principle. A creative system should maintain a persistent problem state, render that state through multiple modules, inspect the products, and use unexpected but useful features to update the problem state. Language, vision, code, simulation, retrieval, planning, experimental design, and evaluation should not function as isolated outputs. They should function as reciprocating systems that modify a shared working-memory-like state. The system should not merely generate a response. It should generate a state change.

The transition from state-spanning coactivity to recursive search-space construction can therefore be summarized in one sequence. First, a set of representations persists across successive states. Second, the retained set defines the current problem space. Third, polyassociative convergence selects a new representation. Fourth, the new representation enters coactivity with the retained set. Fifth, the resulting coalition defines a revised problem space. Sixth, the cycle repeats. Creativity emerges when this recursive sequence produces a novel, useful, and evaluable structure.

The central claim of this section is that recursive search-space construction is the cognitive-level expression of incremental change in state-spanning coactivity. Creative thought is not an escape from ordinary neural dynamics. It is a special use of them. Sustained coactivity preserves the problem. Incremental change modifies it. Polyassociative convergence introduces candidate updates. Novel convergence events create unexpected but relevant additions. Recursive incorporation allows those additions to reshape the next search. Through this process, a mind can construct ideas that were not stored in memory as completed patterns.

5. Insight, Incubation, and Novel Convergence Events

Insight is often experienced as sudden. A solution appears, an analogy clicks, a phrase resolves, a mechanism becomes obvious, or a previously confusing pattern reorganizes into coherence. This suddenness is one reason creativity has often been treated as mysterious. The product enters awareness abruptly, and the prior work that made it possible may be difficult to reconstruct. But the fact that an insight appears suddenly does not mean it was generated instantly.

The present model interprets insight as the conscious appearance of an update that reorganizes the active search space. The crucial event is not simply the arrival of a new representation. New representations enter thought constantly. The crucial event is that the new representation changes how retained contents relate to one another. It makes a previous problem state more coherent, more searchable, or more actionable. In this sense, insight is not merely an answer. It is a restructuring of relevance.

This restructuring can be understood through the concept of a novel convergence event. A novel convergence event occurs when a group of active representations that has not previously been coactive as a group converges on a useful representation, relation, image, or hypothesis. Each element of the coalition may be familiar. Some of the pairwise associations may be familiar. But the full coalition is new, and the update it selects may not have been reachable from any one element alone. The insight is generated by the coalition.

This explains why insights often feel both surprising and appropriate. They are surprising because the active coalition was not part of a familiar pattern. They are appropriate because the update was selected by convergent constraint. The system did not randomly generate novelty. It allowed several partially related contents to jointly search memory and conceptual space. The resulting update can therefore feel as if it was hidden in the problem all along.

A simple association can be remote without being insightful. A remote association becomes an insight when it changes the problem state. If a new idea merely appears and vanishes, it may be interesting but unproductive. If it enters working memory and causes the retained elements to become newly organized, it becomes cognitively powerful. It can transform which features are relevant, which assumptions are questionable, which tests are necessary, and which next thoughts are likely.

This is why insight depends on continuity. A new representation can only reorganize a problem if enough of the problem remains active or potentiated to be reorganized. If the earlier contents have disappeared completely, the new representation has nothing to restructure. If the earlier contents are held too rigidly, the new representation may fail to enter or may be rejected before it can do useful work. Insight requires an intermediate regime in which the problem persists but remains modifiable.

Incubation can be interpreted in the same framework. During incubation, a problem may leave focal awareness, but it does not necessarily return entirely to baseline. Some of its elements may remain potentiated, primed, emotionally tagged, or more easily reactivated. The thinker is no longer deliberately working on the problem, but the system remains biased by it. Later, a new perception, memory, context, or internal association may interact with these residual constraints and produce a useful convergence.

This model explains why incubation can seem passive while still being productive. The active focus of attention may move elsewhere, but the broader short-term store and long-term memory network may remain altered by the unresolved problem. The prior query has not been erased. It has become latent. When a new context supplies a missing element, the earlier problem can be reactivated in a modified state. The resulting insight may feel spontaneous, but it depends on earlier search-space construction.

Incubation also allows constraint relaxation. When a thinker deliberately focuses on a problem, dominant interpretations may remain too active. The same assumptions keep guiding search. Letting the problem recede can reduce the grip of unproductive constraints while preserving weaker traces of the problem. This may allow new representations to enter when the problem is later reactivated. Incubation may therefore work not because thought stops, but because the search space becomes less rigid.

The same process can explain why insights often occur during walks, showers, conversation, reading, dreams, and unrelated tasks. These situations change the active context. They introduce new sensory, bodily, emotional, linguistic, or conceptual material. Some of that material may interact with the latent problem state. A new representation that would not have appeared during direct effort can enter the coalition and reorganize it. The insight occurs when the unresolved problem and the new context converge.

Insight is also closely related to mental imagery. Creative thought often involves rendering a partial problem into an image, scene, diagram, phrase, movement, sound, or simulation. This rendering may add features that were not explicitly specified by the original thought. A person may imagine a mechanism and notice an unanticipated consequence. A sketch may reveal a relation that was not present in the verbal plan. A phrase may suggest a distinction that was not initially intended. These products can then feed back into working memory and alter the search space.

In this way, imagination can function as a search-space expansion system. Higher-order representations specify partial conditions. Lower-order or modality-specific systems render those conditions into richer forms. The rendered product contains details, tensions, and affordances that were not fully contained in the original specification. These details can become new updates. Creative insight may therefore come not only from abstract association, but from the unexpected products of rendering.

This is important for both human and artificial creativity. A creative system should not merely manipulate abstract representations internally. It should render partial ideas into multiple formats, inspect the products, and use emergent features to revise its problem state. In humans, this can happen through mental imagery, gesture, sketching, language, musical improvisation, physical prototyping, and mathematical notation. In AI, it could happen through text generation, diagramming, code execution, simulation, retrieval, image generation, formal proof, and experimental design.

The model also helps explain why insight can be fragile. A novel convergence event may appear briefly and then disappear if it is not stabilized. Many creative people recognize the experience of having an idea that seems important, but losing it before it can be articulated. This happens because the new representation must be incorporated into a larger problem state. It must be connected to retained contents, externalized, rehearsed, written, tested, or otherwise preserved. Without stabilization, the insight may decay before its implications can be explored.

Insight therefore has at least two phases. The first is the convergence event, where a new representation enters the problem state. The second is elaboration, where the system determines what the new representation changes. The first phase may be sudden. The second may be slow. A creative theory, invention, proof, or artwork usually requires both. The flash is not the finished product. It is the beginning of a new search space.

This distinction is especially important in scientific creativity. A scientist may experience a sudden hypothesis, but the hypothesis becomes valuable only if it can generate predictions, explain anomalies, survive objections, and guide tests. The insight is creative because it reorganizes inquiry. It creates a new space of possible experiments, observations, and interpretations. The value of the insight lies not only in the thought itself, but in the sequence of thoughts it makes possible.

The same applies to artificial systems. A model that generates a novel hypothesis has not yet demonstrated scientific creativity in the strong sense. The question is whether the hypothesis reorganizes its subsequent search. Does the system identify new evidence that would matter? Does it distinguish predictions from post hoc explanations? Does it update the hypothesis when tests fail? Does it create new subproblems? Does it preserve the insight across time and use it to guide future inquiry? Without this recursive incorporation, the output remains isolated.

Thus, insight and incubation are not exceptions to the general model. They are special cases of recursive search-space construction. Incubation occurs when elements of an unresolved problem remain latent while the active context changes. A novel convergence event occurs when a newly formed coalition converges on a useful update. Insight occurs when that update reorganizes the retained problem state. Creative work continues when the reorganized state becomes the basis for further search.

The central claim of this section is that creative insight is not simply a sudden answer. It is a state change. A representation becomes insightful when it restructures the active or latent problem space from which subsequent thought will proceed.

6. Hypothesis Generation and Scientific Discovery

Scientific creativity provides one of the clearest examples of recursive search-space construction. A scientific discovery does not usually begin with a fully formed question and a known method for answering it. It often begins with an anomaly, a vague tension, a surprising pattern, an unexplained relationship, a failed prediction, a methodological opportunity, or a dissatisfaction with an existing framework. The first creative act is not solving the problem. It is forming the problem.

A hypothesis is often treated as a candidate answer. But cognitively, a hypothesis is more than an answer. It is a search-space modifier. Once a hypothesis is introduced, it changes what evidence matters, what observations should be reexamined, what experiments should be run, what alternative explanations should be compared, and what new questions become available. The hypothesis reorganizes inquiry.

For this reason, hypothesis generation is not only retrieval. A scientist does not simply search memory for an existing hypothesis. Stored knowledge supplies the materials, but the hypothesis emerges when those materials are placed into a new active configuration. Observations, mechanisms, analogies, constraints, prior findings, uncertainties, and goals are held together long enough for a candidate explanatory structure to form. That structure then changes the next state of thought.

The process can be described as a sequence. First, an unresolved pattern is held active. Second, relevant constraints and background knowledge enter working memory. Third, the active coalition searches for a possible explanatory update. Fourth, a candidate hypothesis enters the coalition. Fifth, that hypothesis changes the search space by implying predictions, tests, and subproblems. Sixth, the results of those tests revise the hypothesis or generate a new one. Scientific thought proceeds by recursively transforming the problem state.

This account helps explain why good hypotheses are not merely novel. A hypothesis may be original but useless if it does not organize inquiry. A productive hypothesis compresses observations, explains anomalies, generates predictions, reveals hidden assumptions, or exposes discriminating tests. Its value lies in how it transforms the search space. It makes new things searchable.

A strong scientific hypothesis does at least five things. It selects relevant evidence from a larger field of possible evidence. It creates expectations about what should be found. It distinguishes itself from alternatives. It suggests methods of falsification or refinement. It generates further questions. In each case, the hypothesis is not merely a content added to thought. It is a structure that changes the direction of future thought.

This is why scientific discovery often depends on problem finding. Many researchers can work on a known problem once it has been clearly named. Fewer can identify the problem that should be named. The creative scientist detects that a current representation of the domain is incomplete. Something does not fit. A pattern is underexplained. A concept is doing too much work. A theory predicts the wrong thing. A measurement opens a new possibility. The mind then constructs a search space around the discrepancy.

Anomaly detection is therefore central to scientific creativity. But an anomaly is not simply a surprising fact. A fact becomes anomalous only relative to a maintained model. The system must preserve expectations long enough for violations to be noticed. It must hold the observed discrepancy in relation to the model that failed. Working memory allows this comparison. It makes the expected and the observed coactive. The anomaly becomes a cognitive object because the mismatch is sustained.

Once sustained, the anomaly can recruit possible explanations. It may activate analogies from other fields, mechanisms from related phenomena, statistical patterns, evolutionary pressures, developmental processes, computational principles, or physical constraints. These elements form a composite problem state. The resulting hypothesis is selected not by one cue, but by the converging influence of the whole coalition. This is multiassociative scientific thought.

Scientific creativity also depends on abstraction. A hypothesis may arise when a thinker represents the current problem at a higher level. Two phenomena that appear unrelated at the surface level may share a deeper structure. This requires the active problem state to drop some details and preserve others. Working memory must perform a selective transformation. It must relax surface constraints while retaining abstract relations. The resulting search space may allow a new analogy or principle to emerge.

Analogy is especially powerful because it imports structure. When a source domain enters the active problem state, it changes which relations are salient in the target domain. A biological system may be understood through an engineering analogy. A cognitive process may be understood through a computational analogy. A disease process may be understood through an ecological analogy. The analogy is not merely decorative. It reorganizes the search space by suggesting mechanisms, variables, tests, and failure modes.

However, analogy must also be constrained. A bad analogy creates misleading search spaces. A good analogy preserves a useful relational structure while allowing irrelevant details to be discarded. Scientific creativity requires the ability to exploit analogies without being captured by them. This again depends on recursive search-space construction. The analogy enters working memory, generates candidate updates, is tested against evidence, and is either refined, restricted, or abandoned.

The same logic applies to experimental design. An experiment is not merely an action taken after a hypothesis is formed. It is a way of constructing a new search space in the world. The experiment creates conditions under which reality can return informative updates. A good experiment does not simply gather more data. It generates data that can change the problem state. It asks the world a question that the current hypothesis makes possible.

In this sense, scientific discovery involves a loop between internal and external search-space construction. Internally, the mind constructs hypotheses, models, and predictions. Externally, it constructs instruments, experiments, measurements, simulations, and datasets. The results of external inquiry return as new representations that update the internal problem state. Discovery occurs when this loop converges on a representation that tracks something real.

This also clarifies the difference between speculation and discovery. Speculation can generate hypotheses internally. Discovery requires that hypotheses be constrained by external evidence. A speculative idea may be creative, but it becomes scientific when it produces predictions, tests, measurements, or explanatory integrations that survive contact with the world. The creative search space must be reality-constrained.

Artificial intelligence systems could, in principle, participate in this process. An AI capable of scientific creativity would need to do more than generate plausible hypotheses from text. It would need to maintain unresolved research problems across time, detect anomalies in relation to expected patterns, generate hypotheses that reorganize inquiry, design tests, interpret results, revise the problem state, and create new tasks from what the results imply. It would need to build search spaces, not merely sample outputs.

Such a system would require persistent memory for unresolved problems. It would need mechanisms for representing open questions, partial explanations, failed attempts, and latent tensions. It would need a working-memory-like state in which observations, hypotheses, constraints, and evaluation criteria could remain coactive. It would need novelty-sensitive gating to decide which anomalies deserve continued attention. It would need external tools for simulation, retrieval, code execution, experiment planning, and data analysis. Most importantly, it would need a recursive loop that incorporates intermediate products into future inquiry.

The distinction between task execution and hypothesis generation is crucial here. A system can execute a research task once the task is specified. It can run a statistical analysis, write code, summarize literature, or test a known hypothesis. But autonomous scientific thought requires the ability to generate the task itself. The system must be able to say: this pattern is unexplained, this contradiction matters, this hypothesis would discriminate between models, this experiment should be performed, this failure reveals a new subproblem.

This is the transition from solving to discovering. Solving operates within a search space that has already been defined. Discovering constructs and revises the search space. A creative scientific system must move between the two. It must form a problem, search for a hypothesis, test the hypothesis, interpret the result, and then reform the problem. Each stage changes the next stage.

Human scientific creativity is limited by working-memory capacity, biological decay, attention, motivation, lifespan, and access to external tools. Artificial systems may eventually exceed some of these limits. They could preserve unresolved problem states indefinitely, maintain many hypotheses in parallel, search across large literatures, run simulations continuously, and revisit earlier ideas after new information becomes available. But these capacities would not automatically produce creativity. They would need to be organized around recursive search-space construction.

The central claim of this section is that hypothesis generation is a specialized form of creativity. It creates not merely a possible answer, but a new structure for inquiry. A scientific hypothesis modifies the search space by determining what evidence matters, what tests should be run, and what new questions can be asked. Scientific discovery occurs when recursive search-space construction becomes disciplined by reality.

7. From Task Completion to Task Invention in AI

Artificial intelligence is moving rapidly from output generation toward task execution. Contemporary systems can summarize documents, write code, operate tools, search literature, generate hypotheses, run analyses, coordinate multi-step workflows, and increasingly participate in scientific research pipelines. This progress is important. It shows that artificial systems can do more than respond with isolated text. They can act across time, decompose goals, use external resources, and produce structured work products.

However, task completion and creative task invention are not the same. A system can execute a task that was given to it without understanding why that task matters, what unresolved problem it serves, or what new task should follow from its result. A system can write a paper from a supplied idea, test a provided hypothesis, or optimize a known benchmark without originating the research direction itself. It may still depend on external prompts to define the relevant search space.

The central transition for creative AI is therefore not merely from short-horizon behavior to long-horizon behavior. It is from assigned-task execution to self-generated task formation. A creative system must be able to construct the problem state that gives rise to the task. It must detect anomalies, maintain unresolved tensions, generate hypotheses, identify missing evidence, formulate tests, interpret failures, and decide what should be investigated next. It must not only answer questions. It must form questions.

This distinction can be expressed in two simplified loops.

The standard agentic loop is: receive goal, plan steps, execute actions, observe results, adjust plan, and report outcome. This loop is powerful, but it begins with an externally supplied goal or problem frame.

The creative cognition loop is different: detect unresolved structure, construct a problem state, generate candidate hypotheses or tasks, test or render them, evaluate the results, revise the problem state, and generate the next task. In this loop, the system does not merely search within a given space. It recursively constructs the space that will be searched next.

The working-memory model developed here suggests that creative AI would need a persistent active state that functions as an internal problem space. This state would contain more than the user prompt. It would include the system’s current goals, unresolved questions, partial hypotheses, failed attempts, relevant evidence, analogies, constraints, uncertainty estimates, evaluative standards, and planned tests. The state would not remain fixed. It would update through partial retention and partial replacement. Some contents would persist because they define the problem. Others would be removed because they no longer help. New contents would enter because they transform the search.

This architecture differs from simply increasing context length. A long context window can store more information, but storage is not the same as active problem-state construction. A creative system must determine which parts of its context are currently operative, which are background, which are unresolved, which are misleading, and which should constrain the next search. The system needs an active working-memory-like state, not merely a large passive record.

It also differs from ordinary chain-of-thought prompting. A chain of reasoning can produce useful intermediate steps, but those steps may remain local to a single task. Creative task invention requires intermediate products to become new problem states. A generated hypothesis should not merely appear in a list. It should alter what the system searches for next. A failed experiment should not merely be reported. It should revise the problem representation. An unexpected result should not merely be noted. It should become a source of new questions.

A creative AI would therefore need several interacting capacities. First, it would need persistent unresolved-problem memory. It must preserve problems that are not solved immediately and allow them to be reactivated later in new contexts. Second, it would need anomaly detection. It must recognize when observations, outputs, or results violate expectations in a way that deserves further inquiry. Third, it would need curiosity-like gating. It must decide which anomalies or possibilities should be held active rather than ignored.

Fourth, it would need a composite working-memory state. It must bring together observations, hypotheses, goals, analogies, constraints, methods, failures, and evaluation criteria in a form that can guide further search. Fifth, it would need multiassociative search. It must generate candidate updates from the whole active coalition rather than from isolated cues. Sixth, it would need recursive incorporation. It must allow useful intermediate products to become part of the next active problem state.

Seventh, it would need multimodal rendering and testing. Creative search-space construction often depends on turning an idea into something more concrete: a sentence, diagram, proof, simulation, codebase, experiment, image, table, design, or physical action. The rendered product exposes constraints that were not available in the abstract formulation. A creative AI should be able to generate such renderings, inspect them, and use their unexpected features to update its problem state.

Eighth, it would need evaluation systems that distinguish productive novelty from noise. These systems would assess coherence, usefulness, predictive power, explanatory depth, novelty, robustness, falsifiability, aesthetic value, implementation feasibility, or goal relevance. Without evaluation, creative search becomes arbitrary generation. Without generation, evaluation has nothing to select. Creative intelligence requires both.

Ninth, it would need task naming. This may be one of the most important capacities. A system becomes self-directed when it can convert an implication into a task. For example: this anomaly should be investigated; this hypothesis requires a discriminating test; this failure suggests a subproblem; this analogy implies a simulation; this missing variable should be measured; this contradiction should be resolved. Task naming is the bridge between internal creativity and external action.

Tenth, it would need memory consolidation. Creative systems should not lose every intermediate product after a session ends. They must store useful problem states, abandoned hypotheses, partial solutions, failed tests, and promising anomalies in forms that can be retrieved and recombined later. Human creativity depends heavily on such unfinished residues. Artificial creativity may depend on persistent libraries of unresolved questions.

This architecture would not eliminate the need for human guidance. In many domains, human values, scientific judgment, ethical constraint, and domain expertise remain essential. But it would shift the role of AI from tool to collaborator. A tool completes a task. A collaborator helps define the task. A creative collaborator notices what is missing, asks what should be asked next, and proposes ways to find out.

The proposed model also helps clarify why current AI systems can appear creative without necessarily being autonomous creators. A system may generate a novel idea when prompted, but the deeper question is whether it can sustain the idea, test it, revise it, remember it, and use it to generate the next line of inquiry. The mark of creative intelligence is not only the production of an impressive output. It is the recursive transformation of the problem state that follows from that output.

This distinction will become increasingly important as AI systems are applied to science. Automated research systems may generate hypotheses, run experiments, and write reports. But scientific creativity requires more than pipeline completion. The system must know when the pipeline has revealed a new question. It must use results to restructure the research space. It must decide whether a failure is a dead end, a measurement problem, a conceptual flaw, or a clue. It must produce not only results, but research directions.

The working-memory model developed here suggests a path toward such systems. Creative AI should be designed around persistent, self-updating problem states. These states should operate as active search spaces that can be recursively modified by intermediate products. The system should maintain continuity without fixation, introduce novelty without incoherence, render ideas into testable forms, and convert useful implications into new tasks.

The central claim of this section is that the next frontier is not merely longer tasks or better answers. It is task invention. A creative artificial intelligence would not only complete the tasks it is given. It would recursively construct the tasks that should be completed.

8. Predictions, Tests, and Conclusion

The claim that creativity is recursive search-space construction should generate empirical and computational predictions. If the model is correct, creative cognition should not appear as a set of independent outputs. It should appear as a temporally extended sequence of partially overlapping problem states. Each state should preserve selected constraints from prior states, incorporate new material, and alter the space of possible next updates.

The first prediction is that creative thought should involve measurable state overlap. During creative cognition, some representations should persist across successive moments while others change. This overlap may be visible in neural population dynamics, representational similarity across time, behavioral traces, verbal protocols, written drafts, sketches, diagrams, code commits, experimental logs, or sequences of intermediate hypotheses. The product may appear sudden, but the process should show continuity.

The second prediction is that creativity should depend on an optimal retention-replacement balance. Too little retention should produce fragmentation. The system fails to preserve the problem long enough for intermediate products to compound. Too much retention should produce fixation. The system remains trapped in a familiar interpretation, schema, or strategy. Creativity should emerge in an intermediate regime where the problem remains active but can be transformed.

The third prediction is that insight should involve reorganization of relevance. An insight should not be detectable only as the addition of a new content. It should change the weighting of retained contents. After insight, previously disconnected elements should become more related, some constraints should become more important, others should become less important, and the next search trajectory should change. Insight is therefore a state transition, not just an answer.

The fourth prediction is that incubation should preserve latent problem structure. During incubation, focal attention may leave the problem, but elements of the problem should remain potentiated, primed, emotionally tagged, or easier to reactivate. Incubation should be most useful when it weakens unproductive constraints while preserving enough unresolved structure for later reactivation. The model predicts that incubation is neither simple unconscious work nor simple rest. It is altered search-space readiness.

The fifth prediction is that creative expertise should involve compressed problem-state construction. Experts should be able to hold high-level representations that stand for many lower-level relations. This allows them to build richer search spaces with fewer active items. However, expertise should also increase the risk of fixation when dominant schemas overconstrain the search. Expert creativity should therefore depend on both structured knowledge and controlled constraint relaxation.

The sixth prediction is that modality shifts should facilitate creativity by reconstructing the search space. Drawing a verbal problem, verbalizing a visual design, simulating a theory, coding an algorithm, building a prototype, or diagramming an argument should create new active features that can feed back into working memory. Creative systems should benefit from cycles of rendering and reinspection because each rendering exposes constraints and affordances not present in the prior formulation.

The seventh prediction is that scientific hypothesis generation should be identifiable as task-generating state transformation. A hypothesis should be considered especially creative when it produces new tests, distinguishes alternatives, reveals new variables, reinterprets anomalies, or opens previously inaccessible subproblems. The value of a hypothesis should be measured not only by plausibility, but by how much it reorganizes inquiry.

The eighth prediction concerns artificial systems. AI systems with persistent, self-updating problem states should outperform systems based on fixed prompts or isolated outputs on tasks requiring open-ended hypothesis generation, autonomous research planning, conceptual synthesis, and long-term scientific discovery. In particular, systems should perform better when they can preserve unresolved questions, render intermediate products, evaluate their implications, and convert those implications into new tasks.

These predictions suggest several kinds of tests. Cognitive neuroscience could examine whether creative tasks show longer or more structured representational overlap than simple recall tasks. Psychology could compare fixation, incubation, and insight in terms of retention-replacement dynamics. Creativity research could analyze drafts, sketches, notes, or protocols as sequences of search-space transformations. AI research could compare architectures that merely generate ideas with architectures that maintain, revise, and act from persistent problem states.

One useful experimental paradigm would compare fixed-query generation with recursive search-space construction. In the fixed-query condition, a model or human participant would be asked to generate many solutions to a problem from a stable prompt. In the recursive condition, intermediate products would be explicitly incorporated into the next problem representation. The model predicts that the recursive condition should produce more coherent, useful, and progressively deepened creative products, especially on tasks requiring hypothesis generation or problem reformulation.

Another paradigm would examine the role of constraint relaxation. Participants or models could be given problems designed to induce fixation. Some conditions would preserve the original framing. Others would force chunk decomposition, analogy, modality shift, or re-rendering. The model predicts that creativity should improve when the active search space is altered without fully destroying continuity with the original problem.

A third paradigm would test AI task invention. Systems could be evaluated not only on whether they solve assigned problems, but on whether they generate meaningful follow-up tasks from partial results. For example, after reading a literature, analyzing a dataset, or running an experiment, the system could be scored on whether it identifies unresolved anomalies, proposes discriminating hypotheses, designs informative tests, and revises its research direction after failure. This would measure creative search-space construction more directly than output novelty.

The model also makes a normative recommendation for AI design. Systems intended for autonomous discovery should be built less like one-shot generators and more like persistent inquiry systems. They should maintain a structured memory of unresolved problems. They should update an active problem state over time. They should use multiple modalities to render and test ideas. They should treat intermediate products as state-changing events. They should generate tasks from implications. They should preserve failures as potentially useful constraints. They should evaluate novelty by its ability to reorganize inquiry.

This approach also suggests a different way of benchmarking creativity. Many creativity benchmarks evaluate isolated outputs. But if creativity is recursive search-space construction, then the process matters. Benchmarks should measure whether a system can improve its own problem formulation, generate increasingly informative hypotheses, use failures productively, shift modalities, and construct better questions over time. The creative trajectory should be evaluated, not only the final artifact.

The conclusion follows directly. Creativity is not merely novelty, free association, divergent generation, or output production. It is the recursive construction of the search space from which useful novelty can emerge. Working memory supplies the active site of this construction. State-spanning coactivity preserves the problem. Incremental change allows the problem to transform. Polyassociative convergence selects candidate updates. Novel convergence events introduce unexpected but relevant material. Recursive incorporation allows intermediate products to reshape the next search.

This account does not require a separate magical faculty of creativity. It treats creativity as a special use of ordinary cognitive mechanisms operating under conditions of novelty, constraint, persistence, and evaluation. The same mechanisms that allow thought to remain continuous also allow it to become original. The same process that lets a mind carry a problem forward also allows the problem to change.

The implications for artificial intelligence are substantial. A system that can complete long-horizon tasks is not necessarily creative. A system that can generate novel outputs is not necessarily creative. A creative system must be able to construct and revise the problem spaces that determine which tasks, hypotheses, and tests should exist. It must be able to move from answering questions to creating questions, and from executing tasks to inventing tasks.

Scientific discovery may depend especially on this capacity. A discovery-oriented intelligence must detect what is unexplained, preserve the tension, construct a hypothesis, test it, revise the problem, and continue. It must build increasingly powerful search spaces around reality. If artificial systems eventually achieve this capacity at scale, they may become capable of autonomous scientific thought.

The deepest claim of this article is therefore simple. Creativity is recursive search-space construction. A mind creates by preserving enough of the old problem to remain coherent, introducing enough new material to transform it, and recursively building the space in which an original answer, hypothesis, task, or discovery can become possible.

References

Amabile, T. M. (1996). Creativity in context: Update to The social psychology of creativity. Westview Press.

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

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.

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. (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.

Boden, M. A. (2004). The creative mind: Myths and mechanisms (2nd ed.). Routledge.

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

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.

Campbell, D. T. (1960). Blind variation and selective retention in creative thought as in other knowledge processes. Psychological Review, 67(6), 380-400.

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.

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.

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.

Gottweis, J., Natarajan, V., et al. (2025). Towards an AI co-scientist. arXiv:2502.18864.

Guilford, J. P. (1950). Creativity. American Psychologist, 5(9), 444-454.

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

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

Kwa, T., West, B., Becker, J., et al. (2025). Measuring AI ability to complete long software tasks. arXiv:2503.14499.

Lu, C., Lu, C., Lange, R. T., Foerster, J., Clune, J., & Ha, D. (2024). The AI Scientist: Towards fully automated open-ended scientific discovery. arXiv:2408.06292.

Mednick, S. A. (1962). The associative basis of the creative process. Psychological Review, 69(3), 220-232.

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

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.

Runco, M. A., & Jaeger, G. J. (2012). The standard definition of creativity. Creativity Research Journal, 24(1), 92-96.

Sawyer, R. K. (2012). Explaining creativity: The science of human innovation (2nd ed.). Oxford University Press.

Simonton, D. K. (1999). Origins of genius: Darwinian perspectives on creativity. Oxford University Press.

Smith, S. M., & Blankenship, S. E. (1991). Incubation and the persistence of fixation in problem solving. American Journal of Psychology, 104(1), 61-87.

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, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.

Yamada, Y., Lange, R. T., Lu, C., Hu, S., Lu, C., Foerster, J., Clune, J., & Ha, D. (2025). The AI Scientist-v2: Workshop-level automated scientific discovery via agentic tree search. arXiv:2504.08066.

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. Advances in Neural Information Processing Systems, 36.