Long-term memory for AI assistants: how to turn Telegram chats into a structured knowledge base

The author shares their experience in developing a long-term memory system for a personal AI assistant that analyzes Telegram correspondence. The primary challenge is the limited context window of large language models, which causes them to quickly 'forget' details from past conversations. A standard approach using RAG (Retrieval-Augmented Generation) and vector search across messages proved insufficient for accurately retrieving context. The article details why the 'naive' chunk-based search method fails to handle user queries regarding specific agreements or budget discussions. The author proposes an alternative architecture designed to transform chaotic chat history into a structured knowledge base. This piece is aimed at developers building intelligent agents and implementing memory systems for LLM applications, offering a practical perspective on overcoming the limitations of modern language models when processing personal data.