EvoArena and EvoMem: A New Paradigm for Memory Evolution of LLM Agents in Dynamic Environments

To address the challenges of deploying in real-world dynamic environments, researchers have introduced the EvoArena benchmark suite and the EvoMem patch-based memory paradigm. Experiments show that current agents have an average accuracy of only 39.6% in dynamic environments, while EvoMem not only improves performance in dynamic environments but also enhances results on standard benchmarks, emphasizing the importance of modeling "evolution" in evaluation and memory.

Large Language Model (LLM) agents perform well in static benchmarks, but real-world environments are constantly changing (e.g., software updates, API changes, evolution of user preferences, etc.). The gap between static evaluation and dynamic reality may lead to misjudgment of agents' true capabilities—agents that are perfect on static benchmarks may fail in dynamic environments.

EvoArena: A New Standard for Dynamic Environment Evaluation

EvoArena is a benchmark suite for evaluating agents' performance in dynamic environments. Its core is modeling environmental changes as a sequence of progressive updates, covering three domains:

• Terminal Environment: Simulate command-line environment changes (addition/removal/modification of commands, file structure changes, etc.)
• Software Environment: Simulate API/software evolution (interface changes, function signature modifications, etc.)
• Social Preference Environment: Simulate social context changes (evolution of user interests, norm updates, etc.)

EvoMem: Patch-Based Memory Paradigm

Traditional memory uses "snapshots" to store environmental states, which has problems such as redundancy, difficulty in tracking changes, and loss of history. EvoMem proposes a patch-based memory approach: instead of snapshots, it records memory evolution, documenting environmental changes as structured update history (including change type, object, content, time, etc.), enabling agents to reason about the process of environmental evolution.

Current Agent Performance

Mainstream agents have an average accuracy of only 39.6% on EvoArena, performing poorly in terminal, software, and social domains, with even worse performance on chained tasks.

Improvement Effects of EvoMem

• Dynamic Environments: Average gain of 1.5% on EvoArena, 3.7% increase in chain-level accuracy
• Standard Benchmarks: 6.1% improvement on GAIA benchmark, 4.8% improvement on LoCoMo benchmark This indicates that EvoMem brings general adaptability and can be transferred to various tasks.

Reasons EvoMem is effective:

1. Enhanced Evidence Capture: Patch-based recording preserves the complete history of environmental evolution, accurately tracks changes, and provides rich context
2. Better Environmental State Representation: It not only records the current state but also includes information on "how it became this way" and "where it might go next" These enable agents to understand change patterns and causal relationships, making more informed decisions.

For Agent Developers

1. Do not ignore dynamic environments; high scores on static benchmarks do not equal reliability in the real world
2. Attach importance to memory system design; patch-based recording may be better
3. Track changes rather than just record states

For Evaluation System Construction

1. Need more dynamic environment benchmarks (like EvoArena)
2. Evaluate both static and dynamic performance simultaneously
3. Test chained tasks to reflect real capabilities

Current Limitations

• Limited domain coverage (only terminal, software, social)
• Relatively simple change patterns
• Patch-based memory increases computational overhead

Future Directions

• Introduce more complex and unpredictable environmental changes
• Study cross-domain adaptability transfer
• Develop agents that actively predict environmental changes
• Explore memory compression to reduce storage overhead

EvoArena and EvoMem provide tools and insights for improving the performance of LLM agents in dynamic environments. Current agents have deficiencies in dealing with environmental changes, and patch-based memory is an effective improvement path.

Core emphasis: Agent system design needs to model "evolution"—both the environment and memory need to evolve. Only systems that can track and understand changes can run reliably in real dynamic environments. EvoArena is a necessary evaluation tool, and EvoMem is a practical architectural reference, which is of great significance for LLM agents to move towards real applications.