SaliMory: Orchestrating Cognitive Memory for Conversational Agents

2026-06-02

AI Summary

This work addresses the challenges faced by existing conversational agents in effectively managing user memory over long-term interactions, where conventional retrieval-augmented approaches degrade reasoning quality and standard reinforcement learning struggles with multi-stage credit assignment. The authors propose an end-to-end language model framework that leverages structured cognitive memory—comprising user facts, preferences, and working memory—to enable selective filtering, integration, and cue-driven recall. A novel hierarchical, stage-wise process reward mechanism combined with reward decomposition through contrastive optimization provides distinct supervisory signals for different memory operations. Experimental results demonstrate that the proposed method reduces memory-related errors by 33%, surpasses state-of-the-art approaches by over 10% in end-to-end accuracy, and more than doubles the rate of high-quality personalized responses.

Abstract

Conversational agents that serve as lifelong companions must maintain persistent memory across all interactions. However, simply expanding context windows with raw retrieval degrades reasoning quality, while training memory agents via standard reinforcement learning creates a severe credit assignment bottleneck in a multi-stage pipeline. To solve this, we introduce SALIMORY, a framework that trains a single language model to manage a cognitively-structured memory-spanning user facts, preferences, and working memory. By introducing a hierarchical stage-wise process reward and reward-decomposed contrastive refinement, SALIMORY provides isolated supervision for distinct memory operations (selective filtering, consolidation, and cue-driven recall) end-to-end. SALIMORY cuts memory-attributed failures by one-third, outperforms the state-of-the-art by over 10% in end-to-end accuracy, and more than doubles the Good Personalization rate.

Problem

• conversational agents
• persistent memory
• credit assignment
• memory management
• reasoning quality

Innovation

• cognitive memory
• process reward
• contrastive refinement
• memory operations
• conversational agents