CoMIC: Collaborative Memory and Insights Circulation for Long-Horizon LLM Agents in Cloud-Edge Systems

2026-05-30

AI Summary

This work addresses the limitations of lightweight large language model (LLM) agents deployed at the edge in long-horizon tasks, where isolated local memory and context inflation hinder persistent memory retention, subgoal tracking, and reflective reasoning. To overcome these challenges, the authors propose the CoMIC framework, which adopts a “centralized reflection, distributed execution” paradigm. Edge agents execute tasks efficiently through hierarchical memory organization and selective historical re-expansion, while a cloud-based LLM asynchronously evaluates execution trajectories to distill reusable experiences and aggregate cross-agent insights via semantic subgoals. Without requiring model parameter updates, this approach enables seamless experience sharing and collaborative reasoning across nodes, significantly improving task progression rates, action grounding accuracy, and task-dependency success for resource-constrained edge agents across five categories of long-horizon tasks.

Abstract

Deploying lightweight Large Language Model (LLM) agents on edge servers can reduce latency and move agentic services closer to users, but resource-constrained edge models often struggle with long-horizon tasks that require persistent memory, subgoal tracking, and reflection. Fine-tuning edge models after deployment is costly and difficult to scale across heterogeneous nodes, while purely local memory leaves agents with isolated experience and growing prompt context. We propose \textsc{CoMIC}, a parameter-update-free cloud-edge framework for Collaborative Memory and Insights Circulation. \textsc{CoMIC} follows a \textit{Centralized Reflection, Decentralized Execution} design: edge agents execute locally using subgoal-oriented hierarchical memory and selective re-expansion of relevant histories, while a cloud-side LLM critic asynchronously evaluates completed trajectories, filters reusable experience, and aggregates cross-agent guidance keyed by semantic subgoal identifiers. Across five long-horizon agent tasks spanning symbolic planning and text interaction, \textsc{CoMIC} improves progress rate and action grounding for weak edge agents and yields task-dependent success-rate gains without updating model parameters.

Problem

• long-horizon tasks
• edge LLM agents
• persistent memory
• subgoal tracking
• experience isolation

Innovation

• Collaborative Memory
• Cloud-Edge Systems
• Long-Horizon Agents
• Parameter-Free Adaptation
• Hierarchical Memory