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

In three linesCoreMem introduces a memory architecture for personalized dialogue agents on edge devices (8 GB VRAM). Replaces cosine similarity with Fisher-Rao metric for retrieval and uses Fisher-guided token distillation for compression. Achieves +4.51 pp gains in open-domain reasoning and +4.17 pp in temporal reasoning on LOCOMO and LongMemEval-S benchmarks.

Summary generated by Claude — human-verified