Hierarchical merging is commonly used to summarize long documents by splitting the input into chunks, summarizing each chunk independently, and merging the results into a final summary. However, because each chunk is processed independently, entity state changes across chunks are not tracked, leading to coherence and consistency issues in the final summary. We propose EntityBB (Entity Blackboard), an agent memory framework that equips chunk-based LLM summarizers with structured entity memory. During sequential chunk processing, the model records and updates entity states through tool calls, while automatic safeguards and model-driven rejection limit noisy writes. Experiments on three long document summarization benchmarks show that EntityBB consistently improves over a controlled hierarchical merging baseline across ROUGE and BERTScore, while NarrativeFactScore indicates improved source-grounded factuality. Targeted human and LLM-as-a-Judge evaluations further indicate fewer entity-level issues and higher judged summary quality.