Purpose: ACE adds a persistent learning loop to AI agents, solving the problems of agents repeating the same mistakes in every session, forgetting effective practices, and ignoring failed experiences. It is based on the ACE paper (Stanford & SambaNova) and Dynamic Cheatsheet.
Key capabilities: It maintains a Skillbook—a persistent collection of strategies that evolves with each task; the learning loop is managed by three specialized roles, among which the Recursive Reflector is the core innovation, writing and executing Python code in a sandbox environment to programmatically search for patterns, locate errors, and iterate until actionable insights are found. All roles are powered by PydanticAI agents with structured output validation, and are routed to 100+ LLM providers through LiteLLM integration, with native support for OpenAI, Anthropic, Google, Bedrock, Groq, and more.
How to use: It offers both interactive installation (recommended) and manual configuration; no fine-tuning, training data, or vector database is required. Strategies can be extracted directly from existing agent logs. ACE is built on a composable pipeline engine, where each step declares requires/provides contracts, uses learning_tail() to compose the standard learning sequence, and also allows custom pipelines; the pipeline engine is framework-agnostic, with immutable context and error isolation.
Applicable scenarios: multi-step agent tasks (such as tau2-bench airline domain tasks, which require tool use and policy adherence), autonomous code translation (ACE + Claude Code translating a library from Python to TypeScript), and more.
Limitations: The materials do not provide specific pricing, cost, or platform support details; the hosted option requires obtaining a demo at kayba.ai.
