Teaching AI Agents to Remember Their Mistakes

What it does ACE wraps existing LLM agents in a feedback loop that records execution traces, extracts strategies from failures, and stores them in a persistent Skillbook. On subsequent runs, the agent retrieves relevant strategies before acting, which the authors claim improves consistency and cuts token use without fine-tuning or vector databases. It ships with ready-made runners for LiteLLM, LangChain, browser-use, and Claude Code, plus a hosted dashboard at kayba.ai.

The interesting bit The Recursive Reflector doesn’t just summarize logs—it writes and executes Python code in a sandbox to programmatically hunt for patterns, isolate errors, and iterate until it finds actionable insights. That turns post-mortem analysis from a narrative exercise into a programmatic debugging session.

Key highlights

• Doubled pass^4 consistency on the Tau2 airline benchmark using 15 strategies learned without reward signals
• Cut browser-automation token costs by 49% over a 10-run learning curve
• Translated roughly 14,000 lines of Python to TypeScript via Claude Code for about $1.50, with zero build errors
• Composable pipeline engine with requires/provides contracts and immutable context
• Can ingest existing agent traces retroactively through the TraceAnalyser runner

Verdict Worth a look if you’re running repetitive agent workflows and tired of watching them hallucinate the same emoji questions. Probably overkill if your agents already nail every task on the first try.