PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents
PsychoAgent separates factual and affective memory, then reranks relevant memories by salience. It retrieved more conflict-critical context, but output-quality differences were not significant.
PsychoAgent filters affective memories by semantic relevance, reranks them by salience, and combines them with factual memory through an executive controller. In **three conflict scenarios**, retrieval scored **0.933**, versus 0.500 and 0.667 for two baselines.
For memory systems where importance is not captured by similarity alone, test a second ranking signal after relevance filtering. Keeping factual and affective stores separate also makes the retrieval policy easier to inspect.
PsychoAgent filters affective memories by semantic relevance, reranks them by salience, and combines them with factual memory through an executive controller. In **three conflict scenarios**, retrieval scored **0.933**, versus 0.500 and 0.667 for two baselines. For memory systems where importance is not captured by similarity alone, test a second ranking signal after relevance filtering. Keeping factual and affective stores separate also makes the retrieval policy easier to inspect. Five blinded raters assessed **27 outputs**. The full system led by **+0.22 SD** after within-rater standardization, but corrected pairwise differences were not statistically significant, and retrieval incurred a small semantic-similarity cost.
This adds affective salience as an inspectable second-stage ranking signal after semantic relevance and supports separating memory types before an executive controller combines them. It extends value-aware retrieval beyond generic utility scores, but the small, statistically inconclusive output study narrows the claim to a promising architecture and retrieval result rather than demonstrated behavioral reliability.