METIS: BRIDGING TEXT AND CODE MEMORY FOR SELF-EVOLVING AGENTS

Self-evolving agents improve by distilling past experiences, but existing systems typically force a choice between storing experiences as natural-language text or as executable code. This paper investigates the trade-offs between these two representations and proposes a hybrid approach to combine their strengths.

The authors conducted a controlled study isolating text memory and code memory on the AppWorld benchmark. They found that text memory is cheaper to construct and more reliable when transferring to new tasks, while code memory offers superior execution efficiency but is costly to build and brittle. Based on these findings, they developed Metis, a system that uses a hierarchical memory structure. Metis stores all experiences as text (categorized into plans, facts, and pitfalls) and only promotes recurring, stable execution plans into validated callable tools.

Key Findings

Metis achieves a better balance between accuracy, efficiency, and construction cost than systems relying on a single representation. Specifically, it improves task accuracy by up to 20.6% over the ReAct baseline while reducing execution costs by up to 22.8%. By treating text as a low-commitment staging layer, Metis avoids the overhead of premature code generation and ensures that only high-value, repeatable procedures are crystallized into tools.

This work provides a systematic understanding of memory representation in agentic systems. By demonstrating that text and code are complementary rather than interchangeable, the authors provide a blueprint for building more efficient and robust self-evolving agents that adapt to open-ended environments without incurring excessive computational or development costs.