Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lemon Agent Technical Report

Haipeng Jiang, Kailong Ren, Zimo Yin, Zhetao Sun, Xin Gan, Guangyi Lv, Ming He, Peng Wang, Congli Yin, Hong Pan, Changwen Zhang, Shan Tong, Zhengyu Xu, Zeping Chen, Yubin Huangfu, Yanzhi Xu, Xing Su, Qin Feng, Dong An, Jianping Fan

TL;DR

The paper addresses inefficiencies in resource use, memory evolution, and multimodal perception in autonomous agents by introducing Lemon Agent, a modular system built on the AgentCortex Planner-Executor-Memory paradigm. It combines a hierarchical self-adaptive scheduling framework, a three-tier progressive context management strategy, and a Self-Evolving Semantic Memory (SES-Memory) with an enhanced MCP toolset to achieve scalable, robust task execution in complex environments. Key contributions include the formalization of the AgentCortex framework, the two-tier scheduling approach, the three-tier context compression, SES-Memory for continual skill extraction, and a strengthened tool suite for perception, search, and geospatial navigation, validated by state-of-the-art GAIA and xbench-DeepSearch results and industrial deployment. The work advances autonomous agent capabilities toward efficient resource utilization, continual knowledge refinement, and high-fidelity perception, with practical impact demonstrated by industrial-scale deployment and open-source availability.

Abstract

Recent advanced LLM-powered agent systems have exhibited their remarkable capabilities in tackling complex, long-horizon tasks. Nevertheless, they still suffer from inherent limitations in resource efficiency, context management, and multimodal perception. Based on these observations, Lemon Agent is introduced, a multi-agent orchestrator-worker system built on a newly proposed AgentCortex framework, which formalizes the classic Planner-Executor-Memory paradigm through an adaptive task execution mechanism. Our system integrates a hierarchical self-adaptive scheduling mechanism that operates at both the overall orchestrator layer and workers layer. This mechanism can dynamically adjust computational intensity based on task complexity. It enables orchestrator to allocate one or more workers for parallel subtask execution, while workers can further improve operational efficiency by invoking tools concurrently. By virtue of this two-tier architecture, the system achieves synergistic balance between global task coordination and local task execution, thereby optimizing resource utilization and task processing efficiency in complex scenarios. To reduce context redundancy and increase information density during parallel steps, we adopt a three-tier progressive context management strategy. To make fuller use of historical information, we propose a self-evolving memory system, which can extract multi-dimensional valid information from all historical experiences to assist in completing similar tasks. Furthermore, we provide an enhanced MCP toolset. Empirical evaluations on authoritative benchmarks demonstrate that our Lemon Agent can achieve a state-of-the-art 91.36% overall accuracy on GAIA and secures the top position on the xbench-DeepSearch leaderboard with a score of 77+.

Lemon Agent Technical Report

TL;DR

The paper addresses inefficiencies in resource use, memory evolution, and multimodal perception in autonomous agents by introducing Lemon Agent, a modular system built on the AgentCortex Planner-Executor-Memory paradigm. It combines a hierarchical self-adaptive scheduling framework, a three-tier progressive context management strategy, and a Self-Evolving Semantic Memory (SES-Memory) with an enhanced MCP toolset to achieve scalable, robust task execution in complex environments. Key contributions include the formalization of the AgentCortex framework, the two-tier scheduling approach, the three-tier context compression, SES-Memory for continual skill extraction, and a strengthened tool suite for perception, search, and geospatial navigation, validated by state-of-the-art GAIA and xbench-DeepSearch results and industrial deployment. The work advances autonomous agent capabilities toward efficient resource utilization, continual knowledge refinement, and high-fidelity perception, with practical impact demonstrated by industrial-scale deployment and open-source availability.

Abstract

Recent advanced LLM-powered agent systems have exhibited their remarkable capabilities in tackling complex, long-horizon tasks. Nevertheless, they still suffer from inherent limitations in resource efficiency, context management, and multimodal perception. Based on these observations, Lemon Agent is introduced, a multi-agent orchestrator-worker system built on a newly proposed AgentCortex framework, which formalizes the classic Planner-Executor-Memory paradigm through an adaptive task execution mechanism. Our system integrates a hierarchical self-adaptive scheduling mechanism that operates at both the overall orchestrator layer and workers layer. This mechanism can dynamically adjust computational intensity based on task complexity. It enables orchestrator to allocate one or more workers for parallel subtask execution, while workers can further improve operational efficiency by invoking tools concurrently. By virtue of this two-tier architecture, the system achieves synergistic balance between global task coordination and local task execution, thereby optimizing resource utilization and task processing efficiency in complex scenarios. To reduce context redundancy and increase information density during parallel steps, we adopt a three-tier progressive context management strategy. To make fuller use of historical information, we propose a self-evolving memory system, which can extract multi-dimensional valid information from all historical experiences to assist in completing similar tasks. Furthermore, we provide an enhanced MCP toolset. Empirical evaluations on authoritative benchmarks demonstrate that our Lemon Agent can achieve a state-of-the-art 91.36% overall accuracy on GAIA and secures the top position on the xbench-DeepSearch leaderboard with a score of 77+.
Paper Structure (20 sections, 1 equation, 4 figures, 1 table, 1 algorithm)

This paper contains 20 sections, 1 equation, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Lemon Agent
  3. AgentCortex Framework
  4. Hierarchical Self-Adaptive Scheduling
  5. Macro-Level Dynamic Orchestrator-workers
  6. Micro-Level Dynamic Tool Parallelization
  7. Hierarchical Context Management Strategy
  8. Self-Evolving Semantic Memory
  9. Tool Augmentation
  10. Intelligent Image Tool
  11. Street View Agent
  12. Multi-Source Search Tool Suite
  13. Task Execution Procedure
  14. Experiments
  15. Experimental Settings
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Overview of the Lemon Agent system. Our orchestration system contains the integration of multi-agent collaboration, sub-worker cluster, self-evolving memory, and tool modules.
  • Figure 2: Comparative performance of Lemon Agent against other well-performed agents across all three difficulty levels of the GAIA benchmark.
  • Figure 3: A case of completing a task by combining useful skills from multiple historical attempts via the SES-Memory module.
  • Figure 4: GAIA cases illustrating the performance of the intelligent image tool and Street View Agent.