Metis: Bridging Text and Code Memory for Self-Evolving Agents

2026-06-23 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Metis is a novel self-evolving agent system that addresses the limitations of existing approaches, which typically rely on either natural-language text or code for experience memory without understanding their trade-offs. A controlled study revealed that text and code memories offer complementary benefits in construction cost, execution efficiency, and transferability, indicating neither is sufficient alone. Guided by these findings, Metis employs a hierarchical dual-representation memory. It organizes textual experiences into execution plans, environment facts, and common pitfalls, while selectively converting frequently reused plans into validated callable tools. This design effectively merges the broad applicability of text memory with the execution efficiency of code memory, incurring tool-generation costs only when justified by repeated reuse. Evaluated on the challenging AppWorld benchmark, Metis improved task accuracy by up to 20.6% over ReAct and reduced execution cost by up to 22.8%, demonstrating a superior balance across accuracy, efficiency, and memory-construction cost.

Key takeaway

For AI Engineers developing self-evolving agents, you should move beyond single-representation memory systems. Implement a hierarchical dual-representation memory like Metis, which combines the broad applicability of text with the efficiency of code. This approach allows you to selectively convert frequently used textual plans into validated callable tools, significantly improving task accuracy by up to 20.6% and reducing execution costs by up to 22.8% compared to traditional methods. Prioritize systems that dynamically balance memory types based on reuse frequency.

Key insights

Self-evolving agents benefit from a hierarchical dual-representation memory that bridges text and code for optimal performance.

Principles

• Text and code memories offer complementary trade-offs.
• Neither memory type alone is sufficient for agents.
• Justify tool generation cost by repeated reuse.

Method

Metis organizes textual experience into execution plans, environment facts, and pitfalls, then selectively crystallizes recurring plans into validated callable tools.

In practice

• Implement dual-representation memory for agent experience.
• Convert frequently used textual plans into callable code tools.
• Benchmark memory systems on cost, efficiency, and transferability.

Topics

• Self-Evolving Agents
• Dual-Representation Memory
• Text Memory
• Code Memory
• Metis System
• Agent Architectures

Best for: Research Scientist, NLP Engineer, AI Scientist, AI Engineer, Machine Learning Engineer

Related on AIssential

• Okay, this unleashed my agent — Self-evolving agents leverage distinct memory and skill management architectures for continuous improvement.
• Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution — ReMe dynamically manages LLM agent procedural memory, enabling smaller models to surpass larger, memoryless counterparts through experience-driven evolution.
• AI Agents of the Week: Papers You Should Know About — Decoupling LLM reasoning from learning via external, hierarchical procedural memory enables efficient, continual agent adaptation.
• AEL: Agent Evolving Learning for Open-Ended Environments — AEL enables LLM agents to learn from experience by dynamically selecting memory retrieval policies and integrating reflective causal insights.
• Metis: A Generalizable and Efficient World-Action Model for Autonomous Driving and Urban Navigation — Decoupling video generation from action prediction in WAMs, using a Mixture-of-Transformers and asymmetric attention, improves efficiency and generalization for autonomous navigation.
• Memory in Agentic AI — Memory is foundational for autonomous AI agents to learn, adapt, and pursue goals effectively over time.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.