AgentBrew: Lifelong Knowledge Brewing from Strong Teachers to Weak LLM Agents

AgentBrew studies a training-free way to transfer interactive knowledge from a strong teacher to a weaker LLM agent through persistent external memory. It has two coupled components: a failure-triggered teacher–Ralph Loop that turns student failures into environment-validated notes, and student-aware synthesis that rewrites teacher knowledge at an executable level for the target student. The paper targets settings with sparse binary feedback and no demonstrations, labels, weight updates, or test-time teacher access. According to the abstract, evaluations and ablations cover coding, math, and tool-use tasks, but the supplied summary does not report quantitative results.