The introduction of LLMs in chatbots has been a revolution in this field with models that are much more capable and intelligent than those that existed previously. Thus, although research and experimentation with these models is rising, there is an aspect, to some extent, forgotten in this research focused mainly on improving their intelligence: the proximity and personalization of these chatbots.
With this objective in mind, in the present project a new chatbot architecture, based on LLMs, with long-term memory over the user's personal data has developed. This development has involved the study of the state of the art of technology related to this topic and the study of human memory to take inspiration from it. Thus, the two vital processes in a memory architecture have been designed: the process of remembering and the process of memorizing.
For the latter, a novel technique using the capabilities of LLMs has been proposed. In order to make use of it, its use has been validated by tests involving volunteers for the human subjective factor and tests specifically designed for LLMs. A new score formula has also been designed to compare the performance of different LLMs in this task.
After analyzing the results, it was concluded that the LLMs are capable of extracting memorable personal data from a conversation in a manner similar enough to that of a person to validate their use. Thus, using the score formula developed, it has also been concluded that the LLM "Llama 2 13b" is the one that best performs this task among the 4 LLMs compared.
Once the validation of the processes has been carried out, the final architecture has been designed and developed involving both processes of memorizing and remembering personal data. At the end of the project a demo of the chatbot was carried out to demonstrate the added value it provides in a conversation.
The final result was a new chatbot capable of memorizing the most important personal data of the user, making use of this data at the most appropriate moments of the conversation. In this way, the chatbot manages to deliver the feeling of closeness, interest in the user and personalization.