This study presents an active memory algorithm that generates responses in generative language models using graph databases. The development of generative language models has picked up pace recently, and there are many commercial services available. However, generative language models are limited by problems such as hallucination, low accuracy and reliability, and limitations in contextualizing and remembering. It is expensive and requires a lot of resources to develop pre-training datasets or fine-tune the base model to address these problems. Instead, well-designed prompts can be used to achieve the desired response, but this requires prompt engineers or training, as well as a thorough understanding of generative language models.All conversations are saved in a graph database to build a memory, and when a user asks a question, it proactively identifies the information it needs and pulls it and its neighbors from the graph database for reference as it generates an answer to the question. This approach streamlines the generation of natural language that disentangles complex and interconnected information in the real world. Research has shown that answering questions based on real-world information increases the efficiency and usability of generative language models in processing information and generating answers.In addition, the memory assist algorithm of the graph database converts various text datasets, not only conversations, into property graph models that can be updated in real time, and provides diverse and accurate information to the generative language model, enabling it to generate accurate responses through diverse information while reducing the size of the language model, thereby increasing efficiency and speed.