OPEN TO DISCUSSION · Making Better Agents

OPEN TO DISCUSSION

Hello world! I am getting deeper and deeper into he Openclaw blackhole and learning a lot of things. some that excite me and some that ruin my mood on a daily basis. I will list them down and I am happy to hear your comments in case you have any suggestions:

CONTEXT: This is the first limit I encountered. Even if my Bot deploys sub agents, eventually its context will get high. In these situations I have it compact he information gathered and move on to a /new session
LLM COOLDOWN (kimi k2): recently my main bot wouldn't go on cool down but the subagents would. This doesnt really make sense and I still haven't found a way to fix it
MEMORY: this is a big one. I often give him instructions that he ALWAYS needs to apply ex:'from now on when I say x you do x'. He does it but when the new session starts he forgets it. So to fix it we created a PERMANENT file in which he puts all the permanent information. Now when I need something to be permanent I just add 'permanent' at the beginning or the end of the prompt. I feel like this kind of works but also that its not the proper way. Everyone I see on youtube just asks their agent to behave a certain way and they remember.
Because I have this memory issue, we started creating different files for him to remember different things. I might need assistance on setting this up.

BONUS QUESTION lol: what local llm would you suggest hat is really good? I was thinking to get deepseek r1 but feel free to make your suggestions!

4 comments