Agentic Context Engine (ACE) adds a persistent learning loop to an agent so that lessons from one run carry into the next. It maintains a Skillbook — a collection of strategies that is updated after tasks, failures, and feedback — and injects those strategies into later runs. There is no fine-tuning, training data, or vector database involved.
Three roles drive the loop. The Agent executes tasks with the Skillbook's strategies attached, the Reflector analyses execution traces to extract what worked and what failed, and the SkillManager curates the Skillbook by adding, refining, and removing strategies. The Recursive Reflector writes and runs Python in a sandbox to search traces for patterns and isolate errors programmatically instead of summarizing them in one pass. All three roles are PydanticAI agents with structured output validation, which routes to many providers through its LiteLLM integration.
Underneath is a composable pipeline engine where each step declares what it requires and what it provides, with immutable context and error isolation; learning_tail() supplies the standard AgentStep -> EvaluateStep -> ReflectStep -> UpdateStep -> DeduplicateStep sequence, and custom pipelines can be assembled from the same parts. Traces recorded elsewhere can be fed in directly with learn_from_traces, without re-running the tasks.
The project reports the following results from its own runs:
| Metric | Result | Context |
|---|---|---|
| Consistency | Doubles pass^4 on the Tau2 airline benchmark | 15 learned strategies, no reward signals |
| Token use | Browser-automation cost cut by 49% | 10-run learning curve |
| Learning cost | ~$1.50 for a 14,000-line Python-to-TypeScript translation | 119 commits, zero build errors, all tests passing |
The approach is based on the ACE paper from Stanford and SambaNova and on Dynamic Cheatsheet.
Features
- Skillbook: a persistent, curated set of strategies that evolves with every task instead of resetting each session
- Three roles: Agent, Reflector, and SkillManager split execution, trace analysis, and strategy curation
- Recursive Reflector: writes and executes sandboxed Python to search traces for patterns and iterate until it finds actionable insights
- Runners: ACELiteLLM for a batteries-included agent, ACE for batch epochs and evaluation, TraceAnalyser for pre-recorded traces, plus browser-use, LangChain, and Claude Code runners
- Pipeline engine: framework-agnostic steps with requires/provides contracts, immutable context, and error isolation
- Learning from existing logs: learn_from_traces extracts strategies from traces you already have
- CLI: ace setup, ace models, ace validate, ace config, an ace-mcp server for editor integration, and a kayba cloud CLI
- Optional extras: browser automation, LangChain, Logfire observability, MCP, and embedding-based skill deduplication
