Memory-Based Attention as a Form of Internal Memory

A large language model is designed for extended dialogues. Observers note that as a conversation gets very long, the model's processing speed per new message does not significantly slow down, and it successfully references information from the beginning of the dialogue. An architectural review confirms the model does not query any external knowledge bases during the conversation. Which of the following best explains this behavior?

LLM Memory Architecture Proposal

Analyzing LLM Memory Mechanisms