SaliMory: Orchestrating Cognitive Memory for Conversational Agents

Researchers introduce SaliMory, a framework that trains language models to manage structured memory for conversational AI agents through hierarchical reward processes and contrastive refinement. The approach reduces memory-related failures by one-third and achieves over 10% improvement in accuracy while doubling personalization rates.

SaliMory addresses a fundamental challenge in building conversational AI systems that can serve as persistent companions: the tension between maintaining rich memory and preserving reasoning quality. Traditional approaches either suffer from degraded performance when expanding context windows with raw retrieval, or face severe credit assignment problems when training memory agents through standard reinforcement learning across multi-stage pipelines. This research contribution matters because conversational AI is rapidly moving toward personalized, long-term interactions where memory coherence directly impacts user experience and trust.

The framework's innovation lies in its hierarchical architecture that cognitively structures memory into distinct layers—user facts, preferences, and working memory—with isolated supervision for specific operations like selective filtering, consolidation, and cue-driven recall. This stage-wise approach with reward decomposition enables more efficient training signals than end-to-end reinforcement learning alone. The measurable improvements—one-third reduction in memory failures and 10% accuracy gains—represent meaningful progress in a field where incremental improvements often require substantial engineering effort.

For the AI industry, this work signals movement toward more sophisticated memory management in language models, directly relevant to companies developing AI assistants, customer service bots, and personalization systems. The doubling of the Good Personalization rate suggests practical applications in real-world deployment scenarios where user satisfaction depends on consistent, accurate memory. Developers implementing conversational agents will likely benefit from memory architectures that separate concerns through explicit supervisory signals. The research indicates that structured approaches to memory management may outperform purely end-to-end learning paradigms.

• →SaliMory reduces memory-related failures by one-third through hierarchical, stage-wise training processes with isolated supervision signals.
• →The framework structures memory into user facts, preferences, and working memory with separate operations for filtering, consolidation, and recall.
• →Performance improvements exceed 10% in end-to-end accuracy with personalization rates more than doubled compared to baseline approaches.
• →The research demonstrates that decomposed reward structures solve credit assignment bottlenecks in multi-stage memory management pipelines.
• →Practical implications extend to deployment of conversational AI assistants requiring persistent, coherent memory across extended interactions.