Layer-Conditional Emotional Weighting: Does Emotion-Aware Memory Consolidation Need to Be Layer-Specific, and Does It Transfer from Dialogue Personalization to Agentic Task Memory?

Abstract

— Personalized LLM agents increasingly rely on structured long-term memory, and a growing line of work conditions memory consolidation and retrieval on a computational emotional-significance signal rather than semantic relevance alone. Existing emotion-aware memory systems, however, treat memory as a single undifferentiated store and are evaluated only on conversational personalization. We ask a narrower, previously untested question: should the influence of emotional significance on memory consolidation and retrieval be conditioned on the target memory layer (e.g., episodic vs. semantic), rather than applied as a single global weight, and do gains from emotion-aware memory observed on dialogue-personalization benchmarks transfer to long-horizon, multi-session agentic task memory? We propose Layer-Conditional Emotional Weighting (LCEW), a memory architecture built on the standard CoALA memory taxonomy in which every consolidation and retrieval weight is generated by a shared, low-parameter hypernetwork conditioned on a learned layer embedding, rather than looked up from an independent per-layer table, so that global-weight scoring (Park et al.) and flat emotion-weighted scoring (Dynamic Affective Memory Management) are recovered as provable limiting cases of a single ridge-regularized objective rather than as separate architectures, and in which memory-layer transitions are modeled as calibrated probabilities rather than hardcoded rules. We specify the full mathematical formulation, algorithm, and a pre-registered experimental protocol comparing LCEW against seven baselines, including a reimplementation of the closest existing system, on both an existing affective-dialogue benchmark and a new agentic-task transfer benchmark. This manuscript is submitted as a Stage 1 Registered Report: no results are reported, and none should be expected at this stage; the object of review is the formal specification, the isolation of layer-conditioning as the untested variable, and the protocol required to test it honestly, including a mandatory replication of a documented personalization-bias risk before any positive claim is made. In-principle acceptance of this Stage 1 manuscript would commit the experiments in Sections XII-XIII to be run and reported, without further alteration to the hypotheses or protocol, in a Stage 2 submission.