PsychoAgent is a cognitive architecture designed to help LLM agents manage memory and decision-making in ways that mirror human cognition, specifically when dealing with unresolved conflict and emotional significance. Rather than relying solely on topical or semantic similarity to retrieve past experiences, the architecture incorporates affective salience—the emotional importance of an event—to ensure that memories relevant to a conflict are accessible during decision-making.
How the Architecture Works
PsychoAgent functions through a conflict-aware executive controller that integrates several components: current context, persona, relationship constraints, and two distinct memory streams.
The system separates memory into factual and affective categories. While factual memories are retrieved based on semantic similarity, affective memories undergo a two-stage process: 1. Semantic Preselection: The system identifies 30 potential memories based on topical relevance. 2. Salience Re-ranking: The system filters these candidates to keep the 10 with the highest affective salience.
This approach allows the agent to prioritize emotionally charged traces that might otherwise be overlooked by standard retrieval methods. Additionally, the architecture includes an "offline recombination" module that simulates a dream-like process, allowing the agent to consolidate recent experiences and reweight memory salience periodically.
Performance in Conflict Scenarios
Researchers tested PsychoAgent against two baselines: a semantic-affective model that lacks salience re-ranking, and a single-memory RAG (Retrieval-Augmented Generation) model. In three controlled conflict scenarios—family financial disputes, workplace criticism, and friendship betrayal—the full PsychoAgent architecture retrieved a significantly higher number of "conflict-critical" memories.
Specifically, PsychoAgent achieved a critical retrieval rate of 0.933, compared to 0.500 for the semantic-affective baseline and 0.667 for the single-memory RAG baseline. This improvement came with a minor cost in overall semantic similarity, suggesting that the architecture successfully prioritizes emotional relevance over pure topical matching.
Human Evaluation and Limitations
Five blinded raters evaluated 27 generated outputs based on persona consistency, memory grounding, conflict sensitivity, naturalness, and action appropriateness. After standardizing the ratings to account for individual rater tendencies, PsychoAgent achieved the highest overall mean score (+0.22 SD).
However, the researchers note that these pairwise differences were not statistically significant after applying corrections. While the architecture shows a favorable descriptive trend in behavioral quality, the study concludes that the evidence supports the architecture as an inspectable mechanism for modeling human-like conflict effects rather than proving definitive behavioral superiority.
Franklin Analysis
The findings suggest that separating factual and affective memory streams is a viable strategy for improving how LLM agents handle complex social interactions. By making the "salience" of a memory an explicit, tunable parameter, the architecture allows for more transparent and inspectable agent behavior. The study’s reliance on a small set of controlled scenarios and a limited number of human raters means that while the mechanism is effective for the tested conflict types, further research would be required to determine its performance across broader, more diverse social contexts.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!