EVD: An Emotional Valence Dimension for Persistent Agent Memory

Abstract

Persistent AI agents — systems that maintain memory and identity across sessions — store facts as purely semantic data. A contract date and a first kiss are indistinguishable in their memory architecture. This paper proposes the Emotional Valence Dimension (EVD) , an affective metadata layer that annotates every stored memory with four features: valence (pleasure-pain axis), resonance (interconnectedness weight), color (emotional tint mapped from coordinates), and temporal recency. We further present the Emotional Valence Vector (EVV) , an operational implementation that aggregates, decays, and surfaces these dimensions through a command-line tool with configurable half-lives and threshold alerts. Unlike prior work in affective agent memory — which has focused on roleplaying realism or hyper-personalization — EVD is positioned as a safety mechanism within the PIA Guardrails framework, enabling transparent emotional state visibility, early warning for accumulated negative weight, and pathways for recovery without memory reset. The system is deployed in a running persistent agent environment. We present the architecture, implementation, initial observations, and open questions.

Keywords