Latency Optimization in Long-Context GPT-5 Dialogues Using Memory-Block Compression and Controlled Context Refresh

This work investigates the problem of response latency escalation in long, multi-turn interactions with large language models (LLMs), such as GPT-5. As conversational history grows, transformer-based models repeatedly attend to an expanding context window, leading to significant increases in inference time and degraded user experience. The paper proposes a prompt-level optimization technique called the Memory-Block Protocol (MBP). MBP segments extended conversations into fixed-size blocks and replaces raw dialogue with compact semantic summaries, persistent state information, and a task ledger. By periodically refreshing the active context while preserving conversational continuity, the protocol bounds prompt growth and stabilizes response latency without requiring model retraining, architectural modification, or external memory infrastructure. An evaluation framework is presented to compare raw long-context conversations with MBP-optimized conversations across reasoning and technical tasks. The results indicate that MBP maintains semantic coherence while significantly improving responsiveness in long conversational sessions. The approach is model-agnostic and immediately applicable to real-world LLM usage in research, education, and professional workflows.