Improving Long-Term Conversational Abilities of Large Language Models Through Behavior Memory

Although Large Language Models (LLMs) have achieved significant results in the field of natural language processing, there are still challenges in handling long-term memory in open domain dialogue systems. Processing lengthy conversation histories requires a significant amount of computational resources, which limits the input window of LLM and makes it unable to effectively handle tasks that require long-term memory. To this end, this article proposes a memory enhancement mechanism based on behavioral memory, which enhances the long-term memory ability of LLM by extracting, storing, and retrieving key behavioral information from conversations through association. This method does not require modifying the internal structure of LLM and is compatible with various open source and closed source models. We conducted experiments on the Generated Virtual Dataset and the results showed that compared to methods based on abstracts and vector databases, the behavioral memory mechanism significantly improved memory retrieval accuracy, response accuracy, and contextual coherence.