PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents

Imagine you're building an AI that needs to remember not just facts but also the emotional weight of experiences. Current models often rely on straightforward factual recall, which can lead to responses that miss the emotional context of a situation. This can result in failures like providing a cold, factual answer when a more empathetic response is needed, which is what's called a lack of affective sensitivity. The challenge is to create a system that can prioritize memories based on their emotional significance while still being relevant to the task at hand. The authors propose PsychoAgent, a cognitive architecture that separates factual and affective memories and uses a conflict-aware executive controller to manage them. This means that when the AI retrieves memories, it first filters them for relevance and then ranks them by their emotional importance, allowing it to respond in a way that feels more human-like. In tests, this architecture outperformed traditional methods in retrieving memories that were critical to resolving conflicts, although it did incur a slight cost in semantic similarity. This work suggests that integrating emotional context into memory retrieval can significantly enhance the performance of LLM agents in real-world applications.

The approach introduces a novel cognitive architecture that integrates affective memory with factual memory.

The evaluation includes controlled scenarios and blinded raters, but lacks significant pairwise differences.