A Motivational Architecture for Conversational AGI

Technical Summary: A Motivational Architecture for Conversational AGI

Problem Statement
Current conversational agents, largely driven by Large Language Models (LLMs), lack an explicit, persistent motivational architecture. While reinforcement learning addresses motivation in physical agents by regulating bodily needs, conversational agents operate in a linguistic regime where the "environment" is a user's evolving mental state and "actions" are speech acts, tool invocations, or silences. Existing systems often rely on implicit motivation via prompts or fine-tuning, or they lack a persistent, need-sensitive state that can be inspected and regulated. The challenge is to design an architecture that balances multiple drives, maintains self-coherence, and adapts to shifting contexts without destabilizing, specifically tailored to the unique sensorimotor loop of dialogue.

Methodology and Framework
The paper proposes a neurosymbolic motivational architecture that reinterprets the OpenPsi lineage (inspired by Bach's MicroPsi) and couples it with the MetaMo formal scaffold. This framework is designed for agents running on a modular execution substrate, which provides persistent symbolic memory, typed skill interfaces, and structured execution traces.

The core architectural innovation is the recasting of homeostasis from bodily deficits to "dialogue-native" needs. The system regulates a vector of needs including competence, uncertainty reduction, affiliation, affinity, legitimacy, nurturing, and aesthetic coherence.

The architecture operates via a ten-stage motivational processing pipeline:

1. Perception: Encoding the dialogue state from user turns, context, and memory.
2. Need Estimation: Measuring deficits relative to target bands and deriving urges.
3. Cognitive Modulation: Automatically deriving modulators (e.g., arousal, valence, dominance) from the need-urge aggregate to set the agent's cognitive posture before situational reasoning.
4. Feeling State: Generating a pre-action affective readout (e.g., concern, curiosity) based on the current need-modulator configuration.
5. Appraisal: Interpreting the specific situation given the established cognitive posture.
6. Candidate Generation: Enumerating typed abstract actions (e.g., "Empathize," "SearchLiterature") using schemas, memory, and reasoning.
7. Scoring and Selection: Evaluating candidates against the goal vector.
8. Action Execution: Dispatching the selected abstract action through the substrate's skills.
9. Outcome Evaluation and Learning: Assessing results against expected need-deltas and updating memory/habits.
10. Governor Blending: Applying the MetaMo update rule to ensure stability and bounded change.

Key Contributions
The paper outlines three primary architectural contributions:

1. Separation of Modulation and Appraisal: The architecture strictly separates cognitive modulation (how the agent is poised) from situational appraisal (how the agent evaluates a specific situation). Modulators are derived from need aggregates to configure the style of cognition (e.g., speed, depth, approach/avoidance), while appraisal interprets the content. This allows the same stimulus to be interpreted differently depending on the agent's current motivational state without collapsing the two into an opaque step.

2. Dual Decision Strategy: The system employs a continuous blend of fast and slow decision paths, analogous to Kahneman's System 1 and System 2.

   • The fast path is urgency-dominant, acting on the most pressing need with minimal deliberation.
   • The slow path is deliberative, evaluating multiple goals and consulting memory.
   • The blend is not a discrete switch but a continuous function modulated by urgency, arousal, and the need for resolution, ensuring the agent can react instantly to crises while still engaging in deep reasoning when appropriate.

3. Architectural Distinction Between Feelings and Emotions: The paper distinguishes between pre-action feelings and post-action emotions.

   • Feelings are pre-action readouts of the need-modulator configuration, serving to pre-configure the decision process and expose the agent's current posture.
   • Emotions are post-action evaluative labels generated after an action is executed and its effects on the need vector are assessed.
   • Keeping these distinct allows each to play its functional role (pre-configuration vs. learning/feedback) without premature unification.

Results and Illustrations
The paper does not present empirical benchmark results but provides two worked examples to demonstrate the framework's application on a modular substrate:

• CompanionAgent: A contemplative dialogue agent optimized for "contemplative attunement." In a scenario involving a distressed user, the agent's high nurturing and affinity needs trigger a "concern" feeling state. The system blends fast empathy with slow guidance, selecting a "Empathize-then-GuideBreath" action. Crucially, the architecture allows the agent to choose silence as a deliberate, motivated action to maximize need regulation, rather than as a system failure.
• ResearchAgent: A tool-rich assistant optimized for "epistemic clarity." When faced with an underspecified query, high uncertainty-reduction needs trigger an "engagement" feeling state. The system prioritizes the slow path to "ClarifyQuestion" before launching tool chains, demonstrating how motivation shapes internal reasoning control (e.g., widening search under high curiosity, tightening filters under high legitimacy).

Significance and Claims
The authors position this work as an architectural synthesis rather than a fully validated mathematical theory or a complete implemented system. The significance lies in:

• Explicit State: Moving beyond "prompt-engineered pseudo-affect" to agents with inspectable, persistent motivational states that causally constrain action selection.
• Scalability: Providing a tractable domain (conversational agents) to develop motivational machinery intended to scale toward Human-Level AGI (HLAGI). The same ingredients—explicit multi-goal regulation, stable dynamics, and typed actions—are designed to extend beyond conversation into broader cognitive settings and social robotics.
• Homeostatic Dialogue: Reinterpreting homeostasis for the linguistic domain, where the agent maintains a dynamic balance of discursive, epistemic, and social needs rather than bodily deficits.

The paper concludes that while unresolved questions remain regarding modulator ontologies and the full mathematical connection to MetaMo's contraction guarantees, this architecture offers a path toward agents that maintain an explicit motivational life structured enough to support memory, tools, self-modification, and social intelligence over time.