luclasm/luclas

Luclas is a Python-based AI agent that you run locally and that improves its behavior over time based on the work you give it. Unlike standard AI assistants that behave the same way every session, Luclas stores what it does and what went right or wrong, and uses that history to inform how it approaches similar tasks in the future. The improvement happens across three mechanisms. The first is experience memory: after every task, a record of what happened is written to a local SQLite database and recalled when similar tasks come up later. The second is a self-updating policy file called core.md, which functions as the agent's operating rules. The agent can rewrite those rules mid-task when it identifies a better approach, and every version of the file is saved so you can compare how the rules evolved. The third is that the database starts completely empty. Everything the agent learns comes from working with you specifically, which means two Luclas instances used for different kinds of work will develop differently. The README is explicit about a specific risk: because the agent writes its own rules, it can develop bad habits that reinforce over time if not corrected. The advice given is to push back explicitly when something goes wrong, read the policy file occasionally, and let the agent fail on real tasks rather than only giving it easy examples. The agent supports a range of tools: it can run shell commands, execute Python code in isolation, read and write files, search the web, and make HTTP requests. Tasks can be scheduled using natural language. It accepts input and delivers results over WeCom, WhatsApp, and Discord, with Telegram and Slack listed as coming. There is also an HTTP API for integrating Luclas into other systems. It connects to any OpenAI-compatible API endpoint, so it works with a local model served by Ollama or similar tools, or with any cloud provider that follows the same format. The project is MIT-licensed.