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

A new self-evolving agent system called Metis combines text and code memory to improve task accuracy by up to 20.6% while cutting execution costs by up to 22.8% compared to the ReAct baseline, according to research submitted on 23 Jun 2026 [1][2]. The system, detailed in a paper posted to arXiv, introduces a hierarchical dual-representation memory that organizes textual experience into execution plans, environment facts, and common pitfalls, then selectively crystallizes recurring plans into validated callable tools [1][2]. This approach is designed to combine the broad applicability of text memory with the execution efficiency of code memory, incurring tool-generation cost only when justified by repeated reuse [2]. The researchers first conducted a controlled study isolating text memory and code memory over an identical set of experiences. They found the two forms exhibit complementary trade-offs in construction cost, execution efficiency, and transferability, concluding that neither representation alone is sufficient [1][2]. Existing self-evolving agents typically commit to one representation at design time rather than deriving the choice from the characteristics of the experience itself [2]. Metis was evaluated on AppWorld, a benchmark for interactive agents [1][2]. The results showed the 20.6% accuracy improvement and 22.8% execution cost reduction over ReAct [1][2]. Compared with other representative self-evolving agent systems, Metis consistently achieved a better balance between accuracy, execution efficiency, and memory-construction cost [1][2]. The work arrives as AI research tools continue to proliferate. Google's NotebookLM, for instance, functions as a retrieval-augmented generation tool that lets users interact with their documents through features such as podcast-like Audio Overviews and AI-generated video summaries, running on Gemini 3.5 models as of June 2026 [5]. The broader AI landscape has also seen platforms like YouTube, owned by Google, expand far beyond video sharing into mobile apps, games, and subscription services, with combined advertising and subscription revenue exceeding $50 billion from Q4 2023 to Q3 2024 [6]. While the Metis paper does not address deployment or product integration, its dual-representation architecture offers a framework for agents that learn from past executions without the rigid design-time choices that have characterized prior systems [1][2].