On the Role of Contextual Information and Ego States in LLM Agent Behavior for Transactional Analysis Dialogues

TL;DR

This paper tackles the deficiency of psychological depth in LLM agents by proposing a TA-inspired Multi-Agent System where each agent comprises Parent, Adult, and Child ego-state modules with separate memories and a memory retrieval mechanism. A central decision process, guided by a life script, selects final responses, and an ablation study compares memory ON versus OFF configurations in a TA dialogue scenario. Results indicate that explicit ego-state modeling yields more nuanced and emotionally grounded interactions, with memory retrieval further enriching non-Adult state activations and social dynamics. The work highlights implications for psychologically grounded simulations in social science, education, and therapy, while noting limitations and avenues for future research (e.g., discounting, stroke concepts, dynamic memory evolution).

Abstract

LLM-powered agents are now used in many areas, from customer support to education, and there is increasing interest in their ability to act more like humans. This includes fields such as social, political, and psychological research, where the goal is to model group dynamics and social behavior. However, current LLM agents often lack the psychological depth and consistency needed to capture the real patterns of human thinking. They usually provide direct or statistically likely answers, but they miss the deeper goals, emotional conflicts, and motivations that drive real human interactions. This paper proposes a Multi-Agent System (MAS) inspired by Transactional Analysis (TA) theory. In the proposed system, each agent is divided into three ego states - Parent, Adult, and Child. The ego states are treated as separate knowledge structures with their own perspectives and reasoning styles. To enrich their response process, they have access to an information retrieval mechanism that allows them to retrieve relevant contextual information from their vector stores. This architecture is evaluated through ablation tests in a simulated dialogue scenario, comparing agents with and without information retrieval. The results are promising and open up new directions for exploring how psychologically grounded structures can enrich agent behavior. The contribution is an agent architecture that integrates Transactional Analysis theory with contextual information retrieval to enhance the realism of LLM-based multi-agent simulations.

Figures (6)

• Figure 1: Key advantages of using LLM-based agents for social simulations, summarized from the analysis Gurcan.
• Figure 2: Conceptual model of the three ego states — Parent, Adult, and Child — and their associated stored knowledge, as described in Transactional Analysis Berne58Stewart12.
• Figure 3: Agent architecture. Each agent consists of three sub-agents — Parent, Adult, and Child — each driven by a distinct prompt and (in the memory-enabled condition) a retrieval-augmented vector memory. At each turn, all sub-agents generate candidate responses based on the current conversational input.
• Figure 4: The prompt defining the life script ($S$) for the agent John. This script guides the agent's decision-making process, shaping its behavior to align with an "I Almost Make It" pattern and the internal conflict of hiding procrastination.
• Figure 5: Distribution of Ego State Selection, Parent (P), Adult (A), and Child (C), for Agents Taylor (left column) and John (right column). The top row shows the results for the Memory ON condition, where agents had access to contextual information from their memory banks. The bottom row shows the results for the Memory OFF condition.