Long-Term Memory for AI Assistants: Transforming Telegram Chats into a Structured Knowledge Base

Large language models typically lack memory, relying on a context window that can handle only a limited number of tokens. Once conversations exceed this limit, the model loses track of previous information, which poses a significant challenge for an assistant designed to retain details about tasks, contacts, and agreements over extended periods. I am developing a personal assistant that analyzes user conversations in Telegram and can respond to queries like, 'What budget did we discuss for the trip to Turkey?' or 'What did we decide regarding the contractor's agreement?'. This requires building a long-term memory system. While the initial approach seemed straightforward—storing all messages in a vector database and performing a top-k search for relevant chunks—I quickly realized this naive retrieval-augmented generation (RAG) method was insufficient. In this article, I will explain why this approach falters and outline the architecture I have implemented as a more effective solution.