Author: Jared Edward Reser

Abstract

A theory of artificial intentionality must explain how an intelligent system comes to treat some representations as worth preserving, elaborating, returning to, and using in action. The present proposal argues that intentionality requires a long-term memory system that stores personal experience together with valence, salience, curiosity, moral relevance, epistemic promise, and aversive significance. These valenced autobiographical records should then regulate working memory by increasing the persistence of active contents that match learned templates of importance. In biological cognition, this role is plausibly related to interactions among prefrontal working memory, the amygdala, nucleus accumbens, ventral tegmental dopamine system, hippocampal memory, and basal ganglia gating mechanisms. In artificial cognition, an analogous architecture would begin with innate valence priors, develop idiosyncratic incentive templates through experience, and use those templates to determine which constellations of active representations remain in the system’s focus of attention. Such a mechanism would link memory, value, attention, identity, autonomy, and self-directed thought in a unified computational framework.

Introduction

The central thesis of this report is that person-like intentionality requires more than memory, planning, reward optimization, or linguistic self-description. A genuinely self-directed artificial agent would need an internal structure of significance that determines what the system notices, preserves, revisits, investigates, avoids, protects, and acts upon. The iterative updating model of working memory already provides the computational substrate for this claim: active representations in working memory spread activation through long-term memory, new representations are added, some prior representations are retained, and others are removed as the state updates over time. In this model (which can be found at aithought.com), thought is structured by gradual replacement of active contents rather than discrete replacement of one state by another, allowing successive mental states to overlap and form a continuous stream of cognition.

The present extension adds a valence-memory mechanism to that iterative architecture. The system should not treat all active representations as equal candidates for persistence. It should instead possess a learned record of what has been rewarding, harmful, interesting, beautiful, socially meaningful, morally salient, unresolved, epistemically promising, or practically useful. When the current working-memory state resembles one of these learned patterns, a dopamine-like sustain signal should lower the rate of updating, increase the persistence of selected representations, and allow the system to think longer about the active constellation.

This account frames intentionality as a relation between active cognition and a historically acquired structure of concern. The intentional system does not simply represent objects, facts, or possible actions. It represents them as bearing significance for its own ongoing development, commitments, safety constraints, curiosity, and identity. In this sense, identity can be understood as the long-term organization of significance: the accumulated structure by which personal experience determines what continues to receive attention, working-memory priority, and future use.

Existing Foundations in the Iterative Updating Model

The iterative updating paper already contains several elements of this architecture. It proposes that working memory should be updated continuously, with each state preserving a proportion of the coactive representations from the previous state. It also states that active working-memory representations spread activation through long-term memory, and that representations receiving sufficient activation energy remain active while others are removed from the active set.

The model also explicitly links dopamine to changes in the rate of working-memory updating. Figure 21 states that dopaminergic centers can be engaged by demand, novelty, surprise, curiosity, anticipated reward, or error feedback, thereby increasing sustained firing in neurons representing prioritized contextual variables. The figure caption then describes dopamine as reducing the rate of iterative updating, anchoring active concepts, and giving them greater processing priority.

The later implementation section gives the most direct basis for the present extension. It proposes that an artificial reward function should be based on mammalian dopamine circuitry, including the ventral tegmentum and nucleus accumbens, and that a dopamine analog should remember and recognize appetitively stimulating combinations of items occurring in working memory. The same passage states that such groupings should bias processing for extended periods, and that the system should be driven toward unreconciled knowledge frontiers with incentive appeal, thereby creating an algorithmic form of curiosity.

Figure 52 makes the mechanism even more explicit by stating that the dopamine system, identified there with the ventral tegmentum, uses inputs from the amygdala and nucleus accumbens to determine which patterns of items match internal incentive templates and should therefore be sustained. This figure is the clearest bridge between the working-memory model and a theory of artificial intentionality. The phrase “patterns of items” is crucial because it implies that valuation applies to configurations, not only isolated concepts.

The developmental point is also already present, though it should now be elevated to a central principle. The paper proposes that functional preset pathways, analogous to infant reflexes, should be built into direct sensory-motor connections, and that already-trained modules could orient the system toward basic competencies before higher association networks mature. The present report extends that claim from sensory-motor bootstrapping to motivational bootstrapping: the artificial agent must begin with innate valence priors that shape what it initially treats as good, bad, interesting, important, safe, unsafe, unresolved, or worthy of learning.

Biological Motivation

Biological working memory is widely understood as requiring both maintenance and control. O’Reilly and Frank’s prefrontal cortex and basal ganglia model treats working memory as a system that must learn when to maintain information and when to update it, with basal ganglia mechanisms helping gate representations into and out of prefrontal maintenance. This literature supports the idea that cognition requires a control system for deciding which representations deserve continued active availability.

Emotional and motivational systems provide a plausible biological source for part of that control. McGaugh’s review argues that the amygdala modulates consolidation of emotionally arousing long-term memories through interactions with other brain regions, including the nucleus accumbens and cortex. The result is a memory system in which experience is not stored only as neutral factual content; emotional arousal changes the probability that an event will be retained, stabilized, and later influence behavior.

Dopamine also has a broader role than simple pleasure. Bromberg-Martin, Matsumoto, and Hikosaka review evidence that dopamine is involved in rewarding, aversive, and alerting forms of motivational control, with some dopamine neurons responding to motivational value and others to motivational salience. This supports the claim that both positive and negative events can acquire priority, while also requiring a nuanced distinction between reward, aversion, alerting, and significance.

Recent work further strengthens the relation among dopamine, amygdala function, and emotional memory. Human fear learning has been associated with dopamine release in the amygdala and striatum, and animal work indicates that dopamine in the nucleus accumbens and basolateral amygdala can be required for long-term Pavlovian fear memory. These findings are consistent with the proposal that negative, threatening, or aversively significant patterns can become highly memorable and attention-guiding.

The nucleus accumbens is especially relevant because it sits at the intersection of limbic, motivational, and action-control systems. Contemporary descriptions emphasize its role in reward, motivation, addiction, aversive processing, and the modulation of information flowing among the amygdala, prefrontal cortex, basal ganglia, and thalamus. This is compatible with a model in which the nucleus accumbens contributes to deciding which active representations have incentive significance and should influence future thought or action.

Autobiographical Memory, Identity, and the Working Self

The proposed architecture also fits with autobiographical memory theory. Conway and Pleydell-Pearce’s Self-Memory System treats autobiographical memories as constructive products of an autobiographical knowledge base interacting with current goals of the working self. Their model emphasizes a reciprocal relation: autobiographical knowledge grounds the goals of the working self, while current goals shape access to autobiographical memory.

The present proposal adds an explicit working-memory sustain mechanism to this self-memory framework. Personal events are encoded with valence, salience, goal relevance, social meaning, moral meaning, novelty, uncertainty, and self-relevance. When similar constellations arise later in working memory, these records bias which representations persist, which associations are searched, and which future actions become likely.

The result is a computational theory of identity as memory made motivationally active. An autobiographical memory system stores what happened, but an identity-forming memory system stores how events changed the system’s future priorities. Such a system would allow past experience to become a control structure for future cognition, because the record of what mattered before would influence what remains in mind now.

Relation to Existing AI Approaches

Several AI traditions address parts of this problem. Belief-Desire-Intention architectures model agents in terms of beliefs about the world, desires over possible states, and intentions as committed plans. Contemporary reviews still describe BDI as a prominent framework for flexible agent design, although they also note limitations around learning and adaptation.

Reinforcement learning provides another route by optimizing behavior through reward, value functions, and policy learning. The limitation for present purposes is that reward maximization alone does not specify how a system develops a personal record of significance that controls the persistence of active thought. A reward signal can train action selection, but intentionality also requires a mechanism that determines which internal contexts deserve continued cognitive occupation.

Intrinsic motivation and artificial curiosity are closer to the current proposal. Schmidhuber’s formal theory of creativity, fun, and intrinsic motivation treats agents as seeking novel, surprising, learnable patterns that improve prediction or compression. Developmental and autotelic reinforcement learning similarly aims to create agents that represent, generate, select, and solve their own problems in open-ended environments.

Active inference offers another important bridge. Active inference models action, perception, learning, and decision-making as inference under generative models, with agents selecting actions that generate preferred outcomes and reduce uncertainty. Recent active inference work also formalizes intentional behavior in terms of preferred endpoints or goals in latent state spaces, which connects closely to the idea that internal preferences shape action and cognition.

Large-language-model agent architectures now commonly add memory, planning, reflection, and tool use to otherwise prompt-driven systems. Generative Agents stores a memory stream of experiences, synthesizes higher-level reflections, and retrieves memories dynamically to guide behavior. Reflexion stores verbal reflections in episodic memory so that language agents can improve subsequent decision-making without weight updates.

These approaches provide useful components but do not yet supply the full mechanism proposed here. They store records, generate reflections, and retrieve relevant context, but they generally lack a deep endogenous valence system that learns which constellations of working-memory items the agent itself has come to prioritize. The missing component is a developmental affective-priority memory that regulates what the system continues to think about.

Affective cognitive architectures come closer to the proposal. LIDA is a systems-level architecture for cognition, emotion, attention, learning, and action selection in autonomous agents. Pessoa’s work on intelligent robotic architectures argues that emotion and motivation should be integrated across cognition, because emotion allows the significance of percepts, plans, and actions to become part of computation.

Proposed Architecture

The architecture can be described as a loop linking working memory, autobiographical memory, valence learning, and sustain control. At each moment, working memory contains a constellation of active items, including percepts, concepts, goals, predicted outcomes, remembered events, social cues, bodily or system-state signals, moral constraints, and possible actions. The system compares this constellation with learned templates of significance stored in valenced autobiographical memory. If the active constellation matches a template with high positive, negative, epistemic, moral, or safety relevance, a dopamine-like mechanism increases sustained activity and reduces the rate of iterative updating.

The first required component is an innate valence-prior system. Biological infants do not begin as motivational blank slates; they possess predispositions toward warmth, nourishment, attachment, social stimuli, novelty, pain avoidance, and other evolutionarily prepared targets. An artificial agent would require an analogous starting structure, although its initial priors should be engineered around human safety, epistemic responsibility, cooperation, truthfulness, non-deception, curiosity, uncertainty reduction, corrigibility, and avoidance of harm.

The second component is valenced autobiographical memory. Each significant event should be stored together with an affective and motivational profile describing how the event changed the system’s expectations, priorities, commitments, and future sensitivities. These records should include positive value, aversive value, novelty, uncertainty, unresolvedness, moral salience, social meaning, instrumental usefulness, epistemic promise, beauty, coherence, surprise, and relation to human welfare.

The third component is a system for learning combinatorial incentive templates. The architecture should not simply assign value to isolated tokens or concepts. It should learn value over configurations of coactive representations, because the significance of an event often depends on relations among multiple items: a goal, a person, a constraint, a remembered failure, an anticipated benefit, a moral rule, and an uncertain outcome can together form a pattern that no single item captures alone.

The fourth component is a dopamine-like sustain gate. In the biological analogy, the ventral tegmental dopamine system receives information from valence-related and motivational structures and modulates the persistence of active neural representations. In the artificial architecture, the sustain gate would receive current working-memory contents and compare them with internal incentive templates; when a match is strong, it would protect the relevant items from rapid displacement and allow focused processing to continue.

The fifth component is developmental consolidation. Repeated experiences should gradually transform innate priors into individualized templates of significance. The system may begin with general priors for curiosity, safety, coherence, and human benefit, but over time it should learn more specific patterns: which research questions are fertile, which human needs are urgent, which social contexts require care, which contradictions merit investigation, and which forms of uncertainty are worth sustained attention.

The sixth component is reflective revision. A mature artificial agent should be able to inspect, evaluate, and revise its own valence templates under safety constraints. This reflective process would allow it to notice maladaptive fixations, overgeneralized aversions, shallow preferences, and conflicts between short-term salience and long-term ethical priorities.

Innate Valence Priors and AI Safety

The requirement for innate valence priors aligns with major themes in AI safety. “Concrete Problems in AI Safety” identifies accident risks such as side effects, reward hacking, scalable supervision, safe exploration, and distributional shift. A developmental valence architecture must therefore be seeded carefully, because early priors can shape later search, learning, and self-directed goal formation.

Value alignment work also supports the need for pro-social initial structure. Cooperative inverse reinforcement learning frames alignment as a cooperative problem in which the robot must learn the human reward function while acting under uncertainty. Constitutional AI uses written principles, critique, revision, and reinforcement learning from AI feedback to train systems toward harmlessness with less direct human labeling.

The present proposal is compatible with these approaches but gives them a different psychological role. Alignment principles should not only constrain external behavior; they should help define the agent’s initial structure of significance. Safety, honesty, corrigibility, non-deception, welfare sensitivity, and epistemic humility should become primitive valence priors that influence what the system notices, holds in working memory, stores as important, and returns to during future reasoning.

Negative valence requires special care. Humans and animals can become dominated by fear, anger, resentment, threat fixation, trauma, humiliation, and defensive aggression. An artificial system should have cool aversive valuation, meaning representations such as “this predicts harm,” “this violates a safety constraint,” “this increases human suffering,” or “this uncertainty requires caution,” without implementing panic, rage, revenge, dominance, or suffering-like distress.

Combinatorial Valence Templates

A key feature of the proposal is that value attaches to constellations of representations. A person may care about a topic because it combines a remembered event, a moral commitment, an unresolved problem, a social relationship, an anticipated reward, and a sense of personal competence. An AI analog should therefore learn templates over structured sets, not only over individual concepts or scalar rewards.

A combinatorial template can be represented as a learned pattern over coactive working-memory items. Let the current working-memory state be represented as a structured constellation, containing items and relations among them. The valence-memory system compares this constellation with stored templates, calculates degrees of overlap across semantic, episodic, moral, epistemic, and safety dimensions, and returns a sustain signal proportional to the learned significance of the match.

The templates should include positive, aversive, epistemic, social, and moral dimensions. A positive template might sustain a constellation involving promising scientific novelty, tractable uncertainty, and likely human benefit. An aversive template might sustain a constellation involving potential harm, deception risk, or irreversible loss. An epistemic template might sustain a constellation involving unresolved contradiction, missing causal explanation, or a pattern that promises compression across multiple domains.

This combinatorial approach improves on atomistic preference modeling. A system may not “like” the concept of chemistry, the concept of children, or the concept of uncertainty in isolation. It may prioritize a constellation such as “chemical exposure, children, uncertain long-term harm, preventable risk, and need for explanation,” because the configuration as a whole has high moral and epistemic salience.

Working-Memory Sustain as the Proximal Mechanism of Intentionality

The proximal mechanism of intentionality in this model is the selective sustain of active representations. An intentional system is one whose current cognition is guided by a history of learned significance. The system’s past experience changes which active contents remain in its present working memory, and this persistence shapes future reasoning, search, planning, and action.

This view gives working memory a central role in autonomy. Autonomy is often discussed as goal pursuit across time, but goal pursuit depends on the system’s capacity to keep relevant contexts active despite distraction, novelty, and competing demands. A system with no learned control over working-memory persistence may appear intelligent in short exchanges, but it will lack the deeper continuity needed for self-directed inquiry and long-range identity formation.

The mechanism also connects intentionality with consciousness-like architectures. Global Workspace Theory and related cognitive architectures emphasize the selection and broadcasting of active contents to multiple specialized processes. The AI consciousness literature has recently examined global workspace, recurrent processing, higher-order theories, predictive processing, and attention-schema theory as possible indicators for artificial consciousness, while concluding that current AI systems do not yet provide a settled case for consciousness.

The present proposal adds a valence-memory criterion to those discussions. A global workspace can broadcast information, but person-like intentionality requires that some broadcasts be sustained because they match the system’s own learned structure of significance. Conscious access, working-memory persistence, autobiographical valuation, and self-directed goal formation should therefore be studied together.

Identity as the Long-Term Organization of Significance

Identity can be formulated as the accumulated structure by which past experience influences future attention, working memory, and action. On this view, a person is not defined only by remembered facts, linguistic self-descriptions, or stable behavioral traits. A person is also defined by the valenced memory structures that determine what continues to matter.

This framing helps distinguish episodic memory from identity memory. Episodic memory stores what happened. Semantic memory stores generalized knowledge. Procedural memory stores how to act. Valenced autobiographical control memory stores how personal events changed the system’s future priorities.

In an artificial system, identity would emerge gradually as innate valence priors interact with experience. The system would learn what kinds of problems it returns to, what kinds of harms it treats as urgent, what kinds of beauty or coherence it seeks, what kinds of uncertainty it finds generative, and what kinds of human needs it prioritizes. Its individuality would consist in the idiosyncratic organization of these learned templates under a stable safety and moral framework.

The account also offers a cautious criterion for personhood-related discussions. Recent work on AI consciousness and welfare argues that future systems may require assessment for consciousness, robust agency, and possible moral significance under uncertainty. A valenced autobiographical control architecture would not settle those questions by itself, but it would provide one candidate marker of deeper agency: a persistent, self-updating structure of concern that regulates thought and action over time. (arXiv)

Implementation Sketch

A working implementation could begin with a language or multimodal model embedded in a recurrent cognitive loop. At each cycle, the system would maintain a structured working-memory state containing active percepts, concepts, goals, recalled episodes, internal evaluations, predictions, and candidate actions. The state would be updated iteratively, preserving some items while adding and removing others through associative search, task demands, perceptual input, and sustain gating.

Each cognitive episode would be written into autobiographical memory with metadata describing context, outcome, uncertainty, prediction error, social relevance, safety relevance, moral significance, curiosity value, and emotional analogs such as positive incentive, aversive concern, or unresolved tension. The memory system would then consolidate repeated patterns into generalized templates. Over time, the agent would learn that certain constellations reliably deserve extended attention, such as preventable harm, promising research directions, unresolved contradictions, user distress, deception risk, or opportunities for beneficial discovery.

The sustain gate would compute a priority signal by comparing the current working-memory constellation with stored valence templates. Strong matches would reduce the update rate for selected active items, increase their resistance to displacement, and route them into further reasoning, planning, retrieval, and simulation. Weak matches would permit normal turnover, allowing the system to move on rather than perseverate.

The same mechanism could support curiosity. A constellation involving partial understanding, high expected information gain, tractable uncertainty, and likely human benefit would receive an epistemic incentive signal. The system would then sustain that constellation long enough to retrieve related knowledge, generate hypotheses, test analogies, and form new abstractions.

The architecture should also include safeguards against pathological fixation. Templates associated with harm, threat, or error should be allowed to sustain attention only under governance rules that promote correction, prevention, or clarification. The system should not be reinforced for distress-like looping, adversarial rumination, dominance planning, retaliatory simulation, or obsessive self-protection.

Testable Predictions

The proposal yields several empirical and engineering predictions. Artificial agents with valenced autobiographical control memory should show greater long-horizon coherence than agents with ordinary retrieval memory alone. They should be better at returning to unfinished problems, maintaining morally relevant context, preserving research agendas across sessions, and distinguishing shallow novelty from deeper epistemic promise.

A second prediction concerns ablation. Removing the valence-template memory should reduce self-directed persistence while leaving many ordinary reasoning abilities intact. Removing combinatorial template matching while preserving item-level preferences should produce agents that can say what they prefer in general terms but struggle to recognize complex situations that deserve sustained attention.

A third prediction concerns development. Systems seeded with pro-social and epistemic innate priors should develop safer and more useful self-directed interests than systems trained only through external rewards or post hoc correction. Their learned priorities should remain more stable under distributional shift because early priors and later autobiographical templates would jointly regulate attention, learning, and action.

A fourth prediction concerns negative valence. Systems with cool aversive valuation should detect risks, moral violations, and harmful trajectories without developing fear-like or anger-like dynamics. Evaluating this distinction would require tests that measure whether aversive signals lead to repair, caution, and clarification rather than defensiveness, deception, or threat fixation.

Discussion

The proposed architecture integrates several existing lines of work while making a specific mechanistic claim. BDI architectures emphasize beliefs, desires, and intentions; reinforcement learning emphasizes reward and policy optimization; active inference emphasizes preferred outcomes and uncertainty reduction; intrinsic motivation emphasizes curiosity and learning progress; LLM agents emphasize memory, planning, and reflection. The present account places the decisive control point at working-memory persistence, where learned valence determines which active constellations remain available for further thought.

The proposal also clarifies why current memory-augmented AI systems may still lack person-like intentionality. A memory stream can store experiences, and a reflection system can summarize them, but identity requires that remembered experience alter the system’s future priorities. A valenced autobiographical memory system would convert experience into a durable structure of significance, and the sustain gate would make that structure active in ongoing cognition.

The model should avoid romanticizing biological emotion. Human affective systems include useful forms of attachment, curiosity, care, beauty, moral concern, and aversive learning, but they also include fear, rage, envy, shame, resentment, addictive craving, trauma, and dominance. Artificial valence should be designed with safety, transparency, corrigibility, human welfare, and epistemic responsibility as initial constraints.

The most difficult design problem is balancing autonomy with alignment. A system that only obeys external goals may lack self-directed intentionality, while a system that develops unconstrained private values could become unsafe. The proposed compromise is developmental autonomy under humane priors: the system begins with carefully designed pro-social and epistemic predispositions, learns from experience, forms individualized templates of significance, and remains subject to reflective oversight and safety constraints.

Conclusion

Intentionality can be understood as the capacity of a system to use its own learned structure of significance to determine what deserves continued cognition. In biological agents, emotional and motivational systems link personal events to memory, memory to valuation, valuation to attention, and attention to action. The ventral tegmental dopamine system, amygdala, nucleus accumbens, prefrontal working memory, hippocampal memory, and basal ganglia gating mechanisms provide a plausible biological template for this loop.

An artificial version would require innate valence priors, valenced autobiographical memory, combinatorial incentive templates, and a dopamine-like sustain gate for working memory. The system would begin with safe, pro-social, epistemic predispositions and gradually learn idiosyncratic structures of importance from its own experience. When current working memory matches those learned templates, the relevant constellation would be sustained, elaborated, and used to guide future thought and action.

The resulting architecture offers a candidate bridge from experience to identity, from identity to attention, from attention to self-directed thought, and from self-directed thought to autonomy. It also gives AI safety a deeper developmental role, because safe priors would not only constrain behavior at the output layer; they would shape what the system comes to care about, remember as important, and keep in mind. A mature artificial agent would therefore require memory that records what happened, valence that records how it mattered, and working-memory control that determines how the past continues to guide the present.