每用户 Docker 容器：如何为每个用户提供自己的 AI 代理

When building Adola, an AI companion on Telegram, the initial architecture involved a single server, one model, a database for user tracking, and clever prompt engineering to maintain conversation separation. However, this setup was abandoned after two weeks due to several issues with shared AI instances. Problems included context contamination, where the model occasionally leaked information between users despite user ID prefixes; memory management challenges, as vector databases are suitable for retrieval but not for the evolving memory required by an AI companion; and other limitations. To address these issues, the author decided to implement separate Docker containers for each user, allowing for better isolation and a more personalized experience.