Large language models process extensive dialogue histories during prolonged interactions without additional memory modules. However, they often overlook or misrecall past information. In this paper, we revisit memory-augmented response generation in the era of LLMs. While prior efforts focus on discarding outdated memories, we argue for their utility in providing contextual cues. These cues aid dialogue systems in understanding the evolution of past events, thereby enhancing response generation. We introduce THEANINE, a framework augmenting LLMs with memory timelines. These timelines illustrate the development and causality of relevant past events. Additionally, we present TeaFarm, a counterfactual-driven question-answering pipeline addressing G-Eval's limitations in long-term conversations.
We will release the data upon acceptance.