Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions