TLDR: EvoMem is a new multi-agent AI framework designed to improve planning capabilities by incorporating a dual-evolving memory system. Inspired by human working memory, it uses a Constraint Memory (CMem) for fixed task rules and a Query-feedback Memory (QMem) for iterative solution refinement. This system, involving a Constraint Extractor, Verifier, and Actor, significantly enhances performance in complex planning tasks like trip scheduling, meeting planning, and calendar management across various large language models.
Artificial intelligence is constantly evolving, and one of its most challenging frontiers is planning – the ability to generate sequences of actions and reason about future states. While large language models (LLMs) have made incredible strides, they often struggle with complex planning tasks, especially those requiring multiple steps or long-term dependencies. A new research paper introduces EvoMem, a multi-agent framework that tackles this challenge by integrating a sophisticated, human-like memory system into the planning process.
EvoMem, short for Evolving Memory, draws inspiration from the working memory model in cognitive psychology. It’s designed to help AI agents coordinate more effectively, track constraints, and correct errors iteratively, which are crucial for successful natural language planning. The framework consists of three specialized agents and two distinct memory modules, working together to refine solutions for complex problems.
The Core Components of EvoMem
At the heart of EvoMem are three agents:
- Constraint Extractor: This agent is responsible for identifying all the rules and limitations from a given task description. For example, in trip planning, it would extract details like total trip duration, specific city visit lengths, or flight restrictions. These extracted constraints are vital for guiding the planning process.
- Verifier: The verifier acts as a quality control agent. It evaluates each proposed solution against the constraints identified by the Constraint Extractor. It provides a score (0-100) indicating how well the plan satisfies the constraints and, crucially, lists any violations or errors. Only plans with a perfect score are accepted.
- Actor: This agent generates the actual solution or plan. It uses the extracted constraints and, in later stages, incorporates feedback from previous attempts to refine its output.
Complementing these agents are two innovative memory modules:
- Constraint Memory (CMem): This module stores the task-specific rules and constraints. It’s designed to evolve across different queries (tasks), meaning it can adapt to new types of problems. However, once a query is initiated, CMem remains fixed throughout the iterative planning process for that specific task, providing a stable anchor for the agents.
- Query-feedback Memory (QMem): This is a dynamic memory that evolves within a single query. It accumulates feedback from the Verifier, recording failed attempts, their scores, and identified errors. This allows the Actor to learn from mistakes and refine its solution over multiple iterations until a valid plan is achieved or a maximum number of attempts is reached.
Both CMem and QMem are reset at the end of each query session, preventing any information from one task from influencing another.
How EvoMem Works in Practice
The planning process with EvoMem unfolds iteratively. First, the Constraint Extractor populates CMem with the task’s rules. Then, the Actor generates an initial solution. The Verifier evaluates this solution. If there are errors, the failed solution and its details are logged in QMem, and the Actor uses this feedback, along with CMem, to generate a refined solution in the next iteration. This cycle continues until a perfect solution is found or the iteration limit is met.
Also Read:
- Structuring AI Decisions for Complex Tasks with ReAcTree
- MemSearcher: Enhancing LLM Search Agents with Smart Memory Management
Demonstrated Effectiveness
The researchers evaluated EvoMem on the NaturalPlan benchmark, which includes diverse tasks such as trip planning, calendar scheduling, and meeting planning. Using Gemini-1.5-Pro as the primary backbone, EvoMem consistently outperformed existing strong baselines, showing significant performance improvements across all tasks. For instance, it achieved an average gain of +11.17% in trip planning. The framework also demonstrated its robustness and generalizability by showing consistent improvements when tested with other LLMs like DeepSeek V3 and GPT-4.1-mini.
Ablation studies further highlighted the critical role of each component, confirming that the dual-evolving memory mechanism and the specialized agents are all essential for EvoMem’s success. The findings suggest that maintaining query-specific memory is sufficient for capturing iterative reasoning and substantially boosting performance in complex planning tasks.
EvoMem represents a significant step forward in multi-agent AI planning, demonstrating the profound impact that well-structured memory systems can have on the capabilities of large language models. For more details, you can read the full research paper here.
