CoreMem: Riemannian retrieval + Fisher distillation for edge AI memory

A team of researchers has introduced CoreMem, a novel memory architecture designed to give personalized dialogue agents persistent, long-term memory while operating on consumer-grade hardware with only 8GB of VRAM. Traditional memory systems rely on isotropic cosine similarity for retrieval and heuristic compression rules, leading to the hubness problem (where certain memories dominate retrieval) and syntactic fragmentation during compression. CoreMem addresses these with two innovations grounded in information geometry: Riemannian retrieval using a locally adaptive Fisher-Rao metric (with Mahalanobis distance and Woodbury acceleration for real-time O(Ndr) search) and Fisher-guided discrete token distillation (FDTD), which hierarchically compresses sentences into tokens using sensitivity scores from Fisher information traces. This approach provides a principled trade-off between compression and KL divergence while preserving structural syntax.

Evaluated on the LOCOMO and LongMemEval-S benchmarks, CoreMem achieved substantial accuracy improvements: +4.51 percentage points in open-domain reasoning and +4.17 in temporal reasoning. The system operates seamlessly within an 8GB VRAM budget, bridging the gap between theoretically grounded lifelong memory and real-world edge deployment. The paper, submitted on arXiv (2606.18406), is 15 pages with 5 figures and targets ACM classes I.2.7 and I.2.6. This work represents a significant step toward making truly persistent, intelligent dialogue agents feasible on devices like smartphones, smart speakers, and edge servers.