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

Original: arXiv:2606.00756v1 Announce Type: new

Abstract:

While placing lightweight Large Language Model (LLM) agents on edge servers can lower latency and bring agentic services nearer to end-users, these resource-limited models frequently falter when handling long-horizon tasks. Such tasks demand persistent memory, the tracking of subgoals, and reflective capabilities. Although fine-tuning edge models post-deployment is an option, it is expensive and hard to scale across diverse nodes. Conversely, relying solely on local memory isolates agents, leading to fragmented experiences and bloated prompt contexts. To address these challenges, we introduce \textsc{CoMIC}, a cloud-edge framework that facilitates Collaborative Memory and Insights Circulation without requiring parameter updates. \textsc{CoMIC} operates on a \textit{Centralized Reflection, Decentralized Execution} architecture. In this setup, edge agents perform local execution using hierarchical, subgoal-oriented memory and selectively re-expand relevant historical data. Meanwhile, a cloud-based LLM critic asynchronously assesses finished trajectories, filters for reusable experiences, and aggregates cross-agent guidance based on semantic subgoal identifiers. Evaluated across five long-horizon agent tasks involving symbolic planning and text interaction, \textsc{CoMIC} enhances both the progress rate and action grounding for weaker edge agents. Furthermore, it delivers task-specific improvements in success rates, all without modifying the underlying model parameters.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC