Cellium Agent
A self-evolving general-purpose AI assistant built on a microkernel architecture, that learns from its own successes and failures through a Bayesian bandit decision loop and can be extended at runtime by describing a new component in natural language.
Cellium Agent is a general assistant that can read, write and search files, write and debug code, search the web and summarize results, process data files into reports and visualizations, and set up scheduled natural-language tasks such as a recurring weather check. Its microkernel combines an event bus, dependency injection and a base tool interface, with a decision-loop control architecture that decides, executes, gets feedback and learns each turn, using a Bayesian bandit to keep improving its choice of actions and a three-layer memory of personality, session and long-term memory with hybrid full-text and vector recall.
Components placed in its components folder are hot-loaded within about three seconds without a restart, run inside a sandboxed process with path mapping and dangerous-method interception, and the agent automatically detects and redacts sensitive information such as private keys, tokens and passwords before writing it out. It supports conversations over a web UI as well as QQ, Telegram, Feishu and WeChat, and installs as a downloadable, pre-packaged binary for Windows, Linux x64/ARM64 and macOS.