Title: SaliMory: Orchestrating Cognitive Memory for Conversational Agents

Abstract:

For conversational agents to function as enduring companions, they require the ability to retain persistent memory throughout every interaction. Yet, a significant challenge arises: merely enlarging context windows through raw retrieval compromises reasoning capabilities, and training memory agents using conventional reinforcement learning introduces a pronounced credit assignment bottleneck within multi-stage pipelines. To address these issues, we present SALIMORY, a novel framework that employs a single language model to oversee a cognitively structured memory system encompassing user facts, preferences, and working memory. SALIMORY achieves end-to-end isolated supervision for specific memory operations—nam selective filtering, consolidation, and cue-driven recall—by implementing a hierarchical stage-wise process reward alongside reward-decomposed contrastive refinement. This approach yields substantial improvements: it reduces memory-attributed failures by one-third, surpasses current state-of-the-art methods by more than 10% in end-to-end accuracy, and more than doubles the rate of effective personalization.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC