Prediction-Driven Conversation Memory: A Dual-Agent Framework for Adaptive Dialogue Summarization

Maintaining coherent long-term memory in open-domain dialogue systems remains a fundamental challenge. Conventional approaches summarize conversation history on fixed-interval or rule-based schedules, which often lead to either redundant computation or stale context. We propose, a dual-agent framework that uses for adaptive memory updates. The framework consists of two cooperating agents: Agent A is a standard dialogue agent that generates responses while Agent B continuously predicts the user's next utterance from three distinct conversational angles. When the user's actual utterance diverges from all predictions, the framework triggers a targeted summary regeneration that incorporates the prediction failure as contextual signal. Agent B's multi-angle prediction strategy---covering topic continuation, critical divergence, and topic shift---enables fine-grained detection of conversational drift. We demonstrate that this prediction-driven paradigm reduces unnecessary summary updates during coherent exchanges while ensuring timely memory refresh when the conversation takes unexpected turns. Experimental results show that NextAttention achieves more contextually relevant summaries with fewer LLM calls compared to fixed-schedule baselines, while maintaining response quality.