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

Researchers introduce CoMIC, a cloud-edge framework that enables lightweight LLM agents on edge servers to handle long-horizon tasks by combining local execution with centralized cloud-based reflection and experience aggregation. The parameter-update-free approach improves performance across symbolic planning and text interaction tasks without requiring model fine-tuning.

CoMIC addresses a critical infrastructure challenge in distributed AI systems: how to deploy capable LLM agents at the network edge while maintaining the reasoning quality typically reserved for cloud-based models. The framework's core innovation lies in its asymmetric architecture—edge nodes handle execution efficiently using hierarchical memory and selective history re-expansion, while a central cloud LLM critic asynchronously evaluates trajectories and synthesizes cross-agent insights. This design sidesteps the traditional friction point of edge AI deployment: the need for expensive model fine-tuning after deployment.

The approach reflects broader industry trends toward edge computing and federated AI systems, driven by latency requirements, privacy concerns, and the need to scale heterogeneous hardware deployments. As edge devices and IoT ecosystems expand, the ability to run agentic workflows locally while leveraging cloud resources for strategic reasoning becomes increasingly valuable. Organizations deploying multi-agent systems across distributed infrastructure face exactly this tension.

For developers and infrastructure teams, CoMIC presents a practical middle ground between fully local models (which suffer isolation and context bloat) and cloud-only deployments (which introduce latency). The parameter-update-free approach is particularly significant—it eliminates operational overhead and versioning complexity common in federated learning scenarios. The experiments demonstrate task-dependent success-rate improvements, suggesting real-world viability across diverse workload types, from planning to interactive tasks.

• →CoMIC enables edge LLM agents to handle long-horizon tasks without parameter updates through collaborative cloud-edge memory sharing.
• →The framework uses centralized reflection with decentralized execution, allowing asynchronous cloud-side trajectory evaluation and cross-agent insight aggregation.
• →Performance improvements on long-horizon tasks demonstrate viability across symbolic planning and text interaction domains.
• →Parameter-update-free design eliminates fine-tuning costs and scaling complexity for heterogeneous edge deployments.
• →Semantic subgoal-keyed experience aggregation enables efficient knowledge transfer across distributed agents.