[D] do you guys actually get agents to learn over time or nah?

Posted by Tight_Scene8900@reddit | LocalLLaMA | View on Reddit | 25 comments

been messing with local agents (ollama + openai-compatible stuff) and I keep hitting the same isue

they don’t really learn across tasks

like:
run something → it works (or fails)
next day → similar task → repeats the same mistake

even if I already fixed it before

I tried different “memory” setups but most of them feel like:

which helps a bit but doesn’t feel like actual learning, more like smarter copy-paste

so I hacked together a small thing locally that sits between the agent and the model:

after a few days it started doing interesting things:

still very janky and probably not the “right” way to do it, but it feels closer to learning from experience vs just retrying prompts

curious what you guys are doing for this

are you:

feels like this part is still kinda unsolved