LOCOMO-CONV Benchmark for Context-Driven Memory Retrieval

LOCOMO-CONV shifts memory evaluation from QA-style probing to in-situ conversational usage across four query styles: dialog, implicit, counterfactual, and composed. The benchmark reveals that high retrieval recall in existing systems does not consistently translate to improved end-to-end response quality during implicit queries.