AdMem: Advanced Memory for Task-solving Agents
AdMem framework integrates semantic, episodic, and procedural memory, enhancing LLM performance in long tasks.
Key Findings
Methodology
The AdMem framework combines short-term and long-term memory through a multi-agent architecture enabling automatic memory generation, reward annotation, and adaptive retrieval. Long-term memory is managed through reward-based evaluation, merging, and pruning, ensuring scalability and continual improvement.
Key Results
- Experiments show AdMem achieves task completion rates from 63.4% to 100% in AgentBoard environments, significantly outperforming ReAct and AWM baselines.
- In tool-query and academic domains, AdMem achieves completion rates of 94.7% and 75.0%, demonstrating adaptability across fields.
- Ablation studies show AdMem improves task completion in Jericho from 45% in the first round to 60% in the third round.
Significance
AdMem addresses the challenge of memory management in long tasks for LLMs, improving task completion rates and robustness. It offers new perspectives for academia and industry, particularly in adaptive intelligent agents and long-term task solving.
Technical Contribution
AdMem provides a novel approach to memory generation and management through a multi-agent architecture and reward-driven memory management. It offers significant advantages in automatic memory generation and adaptive retrieval compared to existing methods.
Novelty
AdMem is the first to integrate semantic, episodic, and procedural memory into a unified framework, addressing shortcomings in existing methods, particularly in long task applications.
Limitations
- In some tasks, memory transfer may cause confusion rather than assistance, affecting performance.
- The complexity of the framework may lead to high computational costs.
- Limited memory management design in online deployments.
Future Work
Future work could explore AdMem's application in more domains and optimize online deployment strategies for memory management to enhance scalability and efficiency in real-world environments.
AI Executive Summary
The AdMem framework integrates semantic, episodic, and procedural memory to address the challenge of memory management in long tasks for LLMs. Existing methods primarily focus on storing factual information, while AdMem achieves automatic memory generation, reward annotation, and adaptive retrieval through a multi-agent architecture, ensuring scalability and continual improvement.
Experimental results show AdMem significantly outperforms existing baselines in multi-task conditions, particularly in tool-query and academic domains. Ablation studies demonstrate AdMem's task completion improvement in Jericho from 45% in the first round to 60% in the third round.
AdMem offers new perspectives for adaptive intelligent agents and long-term task solving, despite some limitations such as memory transfer confusion and high computational costs. Its innovations in memory management provide directions for future research.
Deep Dive
Abstract
Large Language Models (LLMs) show promise as tool-using agents but remain limited in long-horizon tasks that require remembering, organizing, and reusing knowledge. Prior memory approaches aim to resolve the situation, but mainly focus on storing factual information. Recent work on procedural memory improves task reuse, yet often reduces to replaying past successes without addressing failure cases or online scalability. We introduce a unified and automatic memory framework that integrates semantic, episodic, and procedural memory in a bi-level design combining short-term and long-term stores. A multi-agent architecture with actor, memory, and critic agents enables automatic memory generation, reward annotation, and adaptive retrieval. Long-term memory is managed through reward-based evaluation, merging, and pruning, ensuring scalability and continual improvement. Experiments across various environments show that our approach improves robustness and success on long multi-turn tasks compared to existing baselines. This work highlights the importance of comprehensive, adaptive memory for advancing LLM-based agents.