CoreMem: Riemannian Retrieval and Fisher-Guided Distillation for Long-Term Memory in Dialogue Agents

CoreMem introduces a novel edge-cloud memory architecture for personalized dialogue agents, addressing the limitations of existing retrieval methods by implementing Riemannian retrieval using a Fisher-Rao metric and Fisher-guided discrete token distillation (FDTD) for efficient context compression. This approach not only mitigates the hubness problem but also enhances memory retrieval accuracy, achieving improvements of +4.51 percentage points in open-domain reasoning and +4.17 in temporal reasoning on the LOCOMO and LongMemEval-S benchmarks, all while operating within an 8 GB VRAM constraint. The advancements in CoreMem are significant for practitioners developing memory-intensive AI systems, enabling effective long-term memory capabilities on resource-limited hardware.