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

2026-06-16 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

CoreMem is a new resource-efficient edge-cloud memory architecture designed for personalized dialogue agents, addressing severe memory and compute bottlenecks on consumer-grade hardware like 8 GB VRAM edge devices. It unifies its approach through information geometry, replacing traditional isotropic cosine similarity with Riemannian retrieval, which employs a locally adaptive Fisher-Rao metric and Mahalanobis distance with O(Ndr) Woodbury acceleration to penalize "hub memories." Additionally, CoreMem introduces Fisher-guided discrete token distillation (FDTD), a hierarchical sentence-to-token compression mechanism that uses sensitivity scores from Fisher information traces for principled compression. Benchmarked on LOCOMO and LongMemEval-S, CoreMem achieved significant accuracy gains, including +4.51 pp in Open-domain and +4.17 pp in Temporal reasoning, while consistently operating within its strict 8 GB VRAM budget.

Key takeaway

For Machine Learning Engineers deploying personalized dialogue agents on resource-constrained edge devices, CoreMem provides a robust solution for long-term memory. You can achieve significant accuracy improvements in Open-domain and Temporal reasoning, specifically +4.51 pp and +4.17 pp respectively, while adhering to strict 8 GB VRAM budgets. Consider integrating its Riemannian retrieval and Fisher-guided distillation to overcome traditional memory and compute bottlenecks.

Key insights

CoreMem uses information geometry for Riemannian retrieval and Fisher-guided distillation to enable long-term memory on 8 GB VRAM edge devices.

Principles

• Information geometry unifies memory architecture for resource efficiency.
• Fisher-Rao metric penalizes hubness in high-dimensional retrieval.
• Fisher information traces guide principled context compression.

Method

CoreMem employs Riemannian retrieval with a Fisher-Rao metric and O(Ndr) Woodbury acceleration for real-time search. It also uses Fisher-guided discrete token distillation (FDTD) for hierarchical sentence-to-token compression.

In practice

• Deploy long-term memory agents on 8 GB VRAM edge devices.
• Improve Open-domain and Temporal reasoning in dialogue agents.

Topics

• CoreMem
• Dialogue Agents
• Long-Term Memory
• Edge Devices
• Riemannian Retrieval
• Fisher Information
• Context Compression

Best for: AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• MemRefine: LLM-Guided Compression for Long-Term Agent Memory — LLM-guided compression, MemRefine, manages agent memory by using an LLM judge for factual content-based decisions, outperforming surface similarity.
• #1 on memory benchmark LongMemEval with Gemini Flash, not Pro [R] — Cognitive science-informed retrieval architectures can significantly enhance conversational memory performance.
• AriadneMem: Threading the Maze of Lifelong Memory for LLM Agents — AriadneMem is a novel structured memory system designed for long-horizon LLM agents, specifically addressing challenges of "disconnected evidence" and "state updates" within fixed context budgets.
• G-Long: Graph-Enhanced Memory Management for Efficient Long-Term Dialogue Agents — G-Long uses graph-enhanced memory with sLM triplet extraction and attention-aware scoring to boost long-term dialogue consistency and efficiency.
• VikingMem: A Memory Base Management System for Stateful LLM-based Applications — VikingMem introduces a Memory Base paradigm for stateful LLM interactions, improving memory retrieval and generalizability.
• MemoryDocDataSet: A Benchmark for Joint Conversational Memory and Long Document Reasoning — MemoryDocDataSet uniquely benchmarks AI systems on joint conversational memory and long document reasoning, revealing a critical performance gap.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.