People expect Jarvis and get another intern tripping everyday.
That's what has happened with OpenClaw.
I have setup 2 OpenClaw agents Vector DB and some local LLMs, with advanced LLM routing natively. One on Pi 5 and another one on my old Macbook Pro M1.
It will forget it has access to certain tools it build, even if you ask it to contextualize, take snapshot, make SoPs for certain tasks.
I found that I have become a full time support for my OpenClaw rather than the other way around.
When I was able to get more out of it? Event based systems that activates the agent, with specific prompting that makes sure it runs well.
Batching processes, for tasks that need to be done periodically. Again, the most important part will be prompting.
I am right now, trying to formalize event based architecture for it since that is the one that works the best.