Online boundary-aware memory for case-based reasoning agents

Large language model (LLM) agents increasingly operate in streaming case-based reasoning (CBR) settings, where continuous improvement from past experience is crucial. Existing methods achieve this by storing past cases and retrieving similar ones as few-shot examples. This strategy fails near decision boundaries, where highly similar cases have conflicting outcomes and the discriminative factors are not explicit. We introduce Online Boundary-Aware Memory (OBAM), an agent memory architecture that discovers and refines decision-boundary knowledge progressively from the case stream. Unlike offline boundary analysis over a static dataset, OBAM detects boundaries online as the agent encounters contrasting cases and stores structured boundary memory entries encoding shared patterns and discriminative rules. These entries are refined as evidence accumulates, transforming ambiguous case experience into reusable decision knowledge. Across legal, medical, and fraud reasoning tasks, OBAM outperforms in-context learning and state-of-the-art agent memory baselines, demonstrating the value of boundary-aware online memory for continual improvement in streaming CBR agents.