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

Large Language Models (LLMs) such as GPT-5 are widely used in continuous, multi-turn conversational settings by students, professionals, and researchers. However, as conversations progress, the accumulated dialogue history expands the model's effective context, resulting in significant increases in response latency. Users frequently observe delays rising from near-instant output to several minutes in prolonged sessions. This paper analyzes the computational basis of this degradation and proposes a Memory-Block Protocol (MBP) that segments dialogue into manageable blocks, generates compact semantic summaries, extracts stable state variables, and periodically refreshes the active context while remaining within the same chat thread. This approach maintains conversational continuity, reduces redundant token reprocessing, and avoids architectural modification to the underlying model. An experimental evaluation framework is provided to measure latency and coherence across mixed reasoning and technical tasks. The protocol improves responsiveness while preserving semantic fidelity, demonstrating that significant performance optimization can be achieved through structured prompt-level memory compression.