AutoAgent: Evolving Cognition and Elastic Memory Orchestration for Adaptive Agents

Xiaoxing Wang, Ning Liao, Shikun Wei, Chen Tang, Feiyu Xiong

Abstract

Autonomous agent frameworks still struggle to reconcile long-term experiential learning with real-time, context-sensitive decision-making. In practice, this gap appears as static cognition, rigid workflow dependence, and inefficient context usage, which jointly limit adaptability in open-ended and non-stationary environments. To address these limitations, we present AutoAgent, a self-evolving multi-agent framework built on three tightly coupled components: evolving cognition, on-the-fly contextual decision-making, and elastic memory orchestration. At the core of AutoAgent, each agent maintains structured prompt-level cognition over tools, self-capabilities, peer expertise, and task knowledge. During execution, this cognition is combined with live task context to select actions from a unified space that includes tool calls, LLM-based generation, and inter-agent requests. To support efficient long-horizon reasoning, an Elastic Memory Orchestrator dynamically organizes interaction history by preserving raw records, compressing redundant trajectories, and constructing reusable episodic abstractions. These components are integrated through a closed-loop cognitive evolution process that aligns intended actions with observed outcomes to continuously update cognition and expand reusable skills, without external retraining.

Introduction

Despite the rapid advancement of Large Language Models (LLMs), current autonomous agents remain surprisingly brittle. Most rely on static cognition—hand-written prompts and fixed workflows that do not adapt to environmental changes or past failures. If a tool's documentation is outdated or a peer agent behaves unreliably, a standard agent will often repeat the same errors indefinitely.

Furthermore, agents frequently suffer from context bloat. As tasks grow in complexity, interaction histories expand, leading to token redundancy that obscures decision-critical information. AutoAgent addresses these challenges by treating an agent's "mental model" as a dynamic, updatable state rather than a static instruction set.

The AutoAgent Framework

AutoAgent redefines the agent architecture by moving away from rigid reasoning-act cycles toward a more fluid, self-correcting system. The framework is built on three integrated pillars:

Comparison of standard Agent vs. AutoAgent architecture.

1. Evolving Cognition

Instead of fixed system prompts, AutoAgent maintains a structured knowledge base divided into Internal Cognition (understanding of its own tools and skills) and External Cognition (models of peer agents and environmental dynamics). This cognition is not static; it is updated through a "Cognitive Evolution" loop. After completing a task, an LLM-based analyzer compares the agent's original intent with the actual outcome, generating precise text-based updates to refine tool descriptions or synthesize new "composite actions."

2. Contextual Decision-Making

The agent employs a "Select-Execute-Update" cycle, choosing from a unified action space. These actions are categorized into:

  • Emic Actions: Self-reliant problem-solving, such as tool calls or code execution.
  • Etic Actions: Seeking external assistance, such as collaboration with peer agents or human-in-the-loop requests.
Detailed view of Agent Cognition and the Emic/Etic Action Space.

3. Elastic Memory Orchestration (EMO)

To manage long-horizon tasks without losing critical evidence, the Elastic Memory Orchestrator maintains a dual representation of history: lossless raw records and summarized abstracts. A dynamic selector determines which parts of the history require full detail and which can be compressed, significantly reducing token overhead while maintaining reasoning accuracy.

Experimental Evaluation

AutoAgent was evaluated across high-difficulty benchmarks including Retrieval-Augmented Generation (RAG) tasks (HotpotQA, Musique) and tool-augmented environments (GAIA, ALFWorld).

  • Superior Reasoning: On HotpotQA, AutoAgent achieved an accuracy of 0.530, significantly outperforming the IRCoT baseline (0.434).
  • Complex Tool Orchestration: On the GAIA benchmark, AutoAgent reached a success rate of 54.5%, nearly doubling the performance of the strong DeepAgent baseline (27.3%).
  • Memory Efficiency: Ablation studies highlighted the importance of EMO; without it, performance on GAIA dropped from 52.1% to 45.5%, proving that intelligent memory management is as critical as the reasoning engine itself.
  • Robustness in Non-Stationary Environments: In experiments where tools were intentionally made to fail 50% of the time, the self-evolution loop allowed AutoAgent to identify unreliable tools and shift strategies, increasing F1 scores by over 11%.

Conclusion

AutoAgent demonstrates that autonomous agents can bridge the gap between inference and learning without the need for expensive model retraining. By treating cognition as an updatable state and memory as an elastic resource, AutoAgent provides a scalable foundation for agents operating in dynamic, real-world environments. Future work will focus on scaling these capabilities to larger multi-agent societies and incorporating more explicit environmental reward signals.

Summary updated

Abstract

Autonomous agent frameworks still struggle to reconcile long-term experiential learning with real-time, context-sensitive decision-making. In practice, this gap appears as static cognition, rigid workflow dependence, and inefficient context usage, which jointly limit adaptability in open-ended and non-stationary environments. To address these limitations, we present AutoAgent, a self-evolving multi-agent framework built on three tightly coupled components: evolving cognition, on-the-fly contextual decision-making, and elastic memory orchestration. At the core of AutoAgent, each agent maintains structured prompt-level cognition over tools, self-capabilities, peer expertise, and task knowledge. During execution, this cognition is combined with live task context to select actions from a unified space that includes tool calls, LLM-based generation, and inter-agent requests. To support efficient long-horizon reasoning, an Elastic Memory Orchestrator dynamically organizes interaction history by preserving raw records, compressing redundant trajectories, and constructing reusable episodic abstractions, thereby reducing token overhead while retaining decision-critical evidence. These components are integrated through a closed-loop cognitive evolution process that aligns intended actions with observed outcomes to continuously update cognition and expand reusable skills, without external retraining. Empirical results across retrieval-augmented reasoning, tool-augmented agent benchmarks, and embodied task environments show that AutoAgent consistently improves task success, tool-use efficiency, and collaborative robustness over static and memory-augmented baselines. Overall, AutoAgent provides a unified and practical foundation for adaptive autonomous agents that must learn from experience while making reliable context-aware decisions in dynamic environments.
BibTeX
@misc{wang2026autoagent,
  author = {Xiaoxing Wang and Ning Liao and Shikun Wei and Chen Tang and Feiyu Xiong},
  title = {{AutoAgent: Evolving Cognition and Elastic Memory Orchestration for Adaptive Agents}},
  year = {2026},
  eprint = {2603.09716},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.09716},
}
Figure