SaliMory framework gives AI agents persistent memory, cuts failures by 33%

Lifelong conversational agents struggle to maintain persistent memory across interactions. Simply expanding context windows degrades reasoning, while training memory agents with standard reinforcement learning creates a credit assignment bottleneck. Researchers from Amazon (Kai Zhang, Xinyuan Zhang, et al.) propose 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 three distinct memory operations: selective filtering, consolidation, and cue-driven recall. This end-to-end approach eliminates the multi-stage pipeline problem.

Results are striking: SaliMory cuts memory-attributed failures by one-third, outperforms prior state-of-the-art by over 10% in end-to-end accuracy, and more than doubles the Good Personalization rate. The framework is designed for real-world deployment where chatbots must remember user details over months of conversation. By treating memory as a cognitive process rather than a simple retrieval cache, SaliMory sets a new benchmark for personal AI assistants that can truly act as lifelong companions without forgetting past interactions.