Memory is a cornerstone of foundation model-based agents for ultra-long-term interaction. However, hallucinated or noisy information can accumulate inside memory systems and progressively contaminate downstream tasks. Existing systems face two key challenges: (1) memory construction mechanisms trade off between faithful recall and hallucination suppression; and (2) retrieval relies on shallow semantic similarity, failing to distinguish logically irrelevant memories. We propose **MemBook**, a hierarchical time-aware memory framework for hallucination-resistant agents. For memory construction, MemBook introduces a stateful memory stabilization mechanism that prevents hallucinated claims from being consolidated into persistent memory and organizes reliable memories into an evolving Session–Chapter–Memory Point hierarchy. For memory retrieval, MemBook performs reasoning-based hierarchical retrieval over this evolving memory book to identify temporally and logically relevant memories. Experiments on HaluMem-Medium, HaluMem-Long, and LongMemEval show that MemBook improves extraction quality and downstream performance while suppressing cross-stage hallucination propagation. Notably, MemBook remains stable under ultra-long contexts, outperforming GAM by absolute 19.57% QA accuracy improvement on HaluMem-Long.