PsychoAgent: Affect-Sensitive Memory Architecture for LLM Agents
A recent study presents PsychoAgent, an innovative cognitive framework designed for large language model (LLM) agents. This architecture distinguishes between factual and emotional memories, merging them through a conflict-aware executive controller. It processes emotional memories based on semantic relevance and subsequently re-ranks them according to their significance, ensuring that emotionally relevant information is included while maintaining topical coherence. In three controlled conflict scenarios, the complete architecture outperformed both semantic-affective and single-memory RAG baselines in retrieving critical conflict memories (0.933 compared to 0.500 and 0.667), incurring only a minor semantic-similarity cost. Evaluations from five blinded raters on 27 outputs indicated that the full architecture had the highest mean score (+0.22 SD), although corrected pairwise differences lacked significance. The research is accessible on arXiv under identifier 2608.07438.
Key facts
- PsychoAgent is a cognitive architecture for LLM agents.
- It separates factual and affective memory.
- It uses a conflict-aware executive controller.
- Affective memories are filtered by semantic relevance and re-ranked by salience.
- In three controlled conflict scenarios, the full architecture retrieved more conflict-critical memories than baselines.
- Retrieval scores: 0.933 vs. 0.500 and 0.667 for baselines.
- Five blinded raters evaluated 27 outputs.
- Corrected pairwise differences were not significant.
- Paper available on arXiv: 2608.07438.
Entities
Institutions
- arXiv