Improving Agent Memory with Affective Awareness

Summary & Methodology Analysis

PsychoAgent improves agent decision-making by addressing the failure of standard RAG (a technique for providing models with external data by retrieving relevant chunks based on vector similarity) to handle emotional or conflict-heavy context. The architecture uses a conflict-aware executive controller to orchestrate interactions between current state and two distinct memory paths. Factual memory remains tied to standard semantic retrieval, while affective memory uses a novel pipeline involving semantic preselection followed by re-ranking. This re-ranking process uses salience weights to ensure the agent does not overlook critical emotional cues during high-stakes conflict resolution. Periodic offline recombination summarizes events into compact, symbolic text traces to maintain contextual continuity across cycles. The implementation leverages models including Gemini 2.5 Flash, Gemini 2.5 Flash-Lite, and Gemini Embedding 2. Performance data shows the full architecture achieves a critical memory retrieval rate of 0.933, compared to 0.500 for the semantic-affective baseline and 0.667 for the single-memory RAG baseline. Human evaluations resulted in a mean score of +0.22 SD, although the results were not statistically significant. Limitations include the use of hand-authored scenarios, a lack of population-level inference, and potential ceiling effects in evaluations that mask behavioral superiority. Furthermore, the cognitive labels applied in the model act as functional analogies rather than direct representations of biological neural or psychological mechanisms.