arxiv "My agent understands me better": Integrating Dynamic Human-like Memory Recall and Consolidation in LLM-Based Agents