Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EvoMem: Improving Multi-Agent Planning with Dual-Evolving Memory

Wenzhe Fan, Ning Yan, Masood Mortazavi

TL;DR

EvoMem addresses the memory bottleneck in LLM-driven multi-agent planning by introducing a dual-evolving memory system. It couples a fixed Constraint Memory (CMem) with a dynamically updated Query-feedback Memory (QMem) across an iterative Actor-Verifier loop among three specialized agents. Empirical results on the NaturalPlan benchmark show consistent gains across trip planning, calendar scheduling, and meeting planning, achieving state-of-the-art accuracy and demonstrating robustness across multiple backbones. The work highlights memory as a key enabler for effective planning in LLM-based multi-agent systems and points to future work on long-term workflow memories to support more complex reasoning tasks.

Abstract

Planning has been a cornerstone of artificial intelligence for solving complex problems, and recent progress in LLM-based multi-agent frameworks have begun to extend this capability. However, the role of human-like memory within these frameworks remains largely unexplored. Understanding how agents coordinate through memory is critical for natural language planning, where iterative reasoning, constraint tracking, and error correction drive the success. Inspired by working memory model in cognitive psychology, we present EvoMem, a multi-agent framework built on a dual-evolving memory mechanism. The framework consists of three agents (Constraint Extractor, Verifier, and Actor) and two memory modules: Constraint Memory (CMem), which evolves across queries by storing task-specific rules and constraints while remains fixed within a query, and Query-feedback Memory (QMem), which evolves within a query by accumulating feedback across iterations for solution refinement. Both memory modules are reset at the end of each query session. Evaluations on trip planning, meeting planning, and calendar scheduling show consistent performance improvements, highlighting the effectiveness of EvoMem. This success underscores the importance of memory in enhancing multi-agent planning.

EvoMem: Improving Multi-Agent Planning with Dual-Evolving Memory

TL;DR

EvoMem addresses the memory bottleneck in LLM-driven multi-agent planning by introducing a dual-evolving memory system. It couples a fixed Constraint Memory (CMem) with a dynamically updated Query-feedback Memory (QMem) across an iterative Actor-Verifier loop among three specialized agents. Empirical results on the NaturalPlan benchmark show consistent gains across trip planning, calendar scheduling, and meeting planning, achieving state-of-the-art accuracy and demonstrating robustness across multiple backbones. The work highlights memory as a key enabler for effective planning in LLM-based multi-agent systems and points to future work on long-term workflow memories to support more complex reasoning tasks.

Abstract

Planning has been a cornerstone of artificial intelligence for solving complex problems, and recent progress in LLM-based multi-agent frameworks have begun to extend this capability. However, the role of human-like memory within these frameworks remains largely unexplored. Understanding how agents coordinate through memory is critical for natural language planning, where iterative reasoning, constraint tracking, and error correction drive the success. Inspired by working memory model in cognitive psychology, we present EvoMem, a multi-agent framework built on a dual-evolving memory mechanism. The framework consists of three agents (Constraint Extractor, Verifier, and Actor) and two memory modules: Constraint Memory (CMem), which evolves across queries by storing task-specific rules and constraints while remains fixed within a query, and Query-feedback Memory (QMem), which evolves within a query by accumulating feedback across iterations for solution refinement. Both memory modules are reset at the end of each query session. Evaluations on trip planning, meeting planning, and calendar scheduling show consistent performance improvements, highlighting the effectiveness of EvoMem. This success underscores the importance of memory in enhancing multi-agent planning.

Paper Structure

This paper contains 14 sections, 3 figures, 8 tables.

Table of Contents

  1. Introduction
  2. EvoMem
  3. Multi-Agent Design
  4. Memory Modules
  5. Proposed Framework
  6. Experiment
  7. Main Result
  8. Ablation Study
  9. Conclusion and Future work
  10. Related Work
  11. Working Memory Model in Cognitive Psychology
  12. Ablation: Effect of Self-Correct Rounds
  13. Details on LLM Agents
  14. Examples of Different Agents in EvoMem

Figures (3)

  • Figure 1: The dual-evolving memory mechanism in EvoMem: CMem evolves across queries to establish fixed, task-specific constraints for the actor and verifier. QMem evolves within a single query by recording failed attempts (solution, score, errors), providing iterative feedback until the solution is verified or the turn limit is reached.
  • Figure 2: (a) Performance comparison of EvoMem against other baseline methods. (b) Performance comparison of EvoMem gradually adding one more component. (All experiments are conducted with Gemini-1.5-Pro.)
  • Figure 3: Working memory model proposed in 1974 baddeley2012working.