LLM Agent Adaptation via Semantic and Episodic Memory

A new memory-augmented framework enables large language model (LLM) agents to learn classification tasks from labeled examples without parameter updates. The approach stores instance-level critiques in episodic memory and distills them into reusable task-level guidance in semantic memory. Across diverse tasks and models, the best self-critique strategy using both memory types achieved an average improvement of 8.1 percentage points over zero-shot baselines and 4.6 points over a RAG-based baseline relying only on labels. Performance varied substantially across models and domains, leading the researchers to introduce 'suggestibility' as a metric to explain this variation. The paper is available on arXiv under identifier 2510.19897.