PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents
Mohammad Amanlou, Parham Abed Azad, Farbod Davoodi, Mostafa Masumi, Behnam Bahrak, Abdol-Hossein Vahabie
Why It Matters
What makes this one worth your time
This research could lead to more human-like LLM agents capable of better handling emotionally charged contexts, which is crucial for applications in mental health, education, and interactive AI.
PsychoAgent enhances LLM memory retrieval by integrating affective significance and conflict awareness.
Summary
The paper presents PsychoAgent, a cognitive architecture for LLM agents that distinguishes between factual and affective memory, integrating them through a conflict-aware executive controller to enhance memory retrieval in conflict scenarios.
Key contributions
- Introduction of a cognitive architecture that separates factual and affective memory.
- Demonstration of improved memory retrieval in conflict scenarios compared to existing baselines.
- Evaluation of the architecture's performance through blinded assessments of output quality.
Notable insights
- The architecture's ability to filter and re-rank affective memories based on salience is a novel approach to memory retrieval in LLMs.
- The use of a conflict-aware executive controller to manage memory retrieval is an innovative step towards more sophisticated cognitive architectures.
Possible limitations
- The significance of the pairwise differences in evaluation results was not established, indicating potential issues with statistical power or evaluation methodology.
- Not stated in the abstract.
Abstract
arXiv:2608.07438v1 Announce Type: new Abstract: Human-like cognition does not select past experience by topical similarity alone: affective significance and unresolved conflict also shape what becomes accessible. We present PsychoAgent, a cognitive architecture for LLM agents that separates factual and affective memory and integrates both through a conflict-aware executive controller. Affective memories are first filtered by semantic relevance and then re-ranked by salience, preserving topical fit while allowing emotionally important traces to enter the prompt. Across three controlled conflict scenarios, the full architecture retrieved more conflict-critical memories than semantic-affective and single-memory RAG baselines (0.933 vs. 0.500 and 0.667), with a small semantic-similarity cost. Five blinded raters evaluated 27 outputs. After within-rater standardization, the full architecture had the highest overall mean (+0.22 SD), but corrected pairwise differences were not significant. A three-day illustrative trace further shows persistent affect, offline memory recombination, and selective memory reweighting. The findings support affect-sensitive retrieval as an inspectable mechanism for modeling human-like conflict effects in LLM agents.