[IDEA] Adding a self-improvement step?

[flatpackfan]

First off, I absolutely love this project. Use it daily, and it creates amazing results for me. Great work and thanks for putting your time and effort into it.

Last week I came across Hermes Agent and I think it might have something that could potentially be useful for this project as well. They claim that their agent is self-improving / learning. Which makes a lot of sense.

What if Agent Hive would, an additional agent that can review the work done, very similar to the Hygenic Review, but rather focus on capturing the learnings in a (long-term) memory system. The most obvious step would be to extract patterns into skill.md files. There are plenty of skills out there that help with writing good skills already so the agent could utilize that.

[tctinh]

Hi @flatpackfan, thanks for the idea.
First I would want to clarify about "self-improvement" and "memory for agent" matter, the reason why I'm not very keen of memory stuff reference #17
Back then I was a part of R&D teams in my organization, we did read plenty of self improvement paper also try a bunch of these framework & memory tools as well, but in the end it not working as expected, the issue is context window and focus lost, agent can only improvement if they actually remember stuff, the way they "remember" at the moment is just rely on context window but big context window is a cure and a poison at the same time, if context window too big model will lose focus and eventually forgot stuff. Let wait till someone make a breakthrough on the memory matter.

For now Hive direction would be smart context window management because despite memory didn't work as good as we wanted, smart managed context window is working and it work really well, as the "AGENTS.md" serve as psudo-memory it should evolve as project continue would be great way to do memory at the moment.

That would solve the memory part the next part is the knowledge which I really like the idea that having a seperate agent to do the "remember" part but introduce seperated agent would cause duplicate with Hygenic Reviewer hence cause token bloat same problem as Oh My Opencode which I want to avoid. Also on the agent create skill matter too much skill would degrade agent performance greatly so I would like to avoid that as well as current skill set of agent is cherry pick and reprompt from popular repo Superpowers if agent come up with their own skill quality will not as good as well, alternative would be the context that agent already produce when hiving those can be condensed into core project knowledge.

I'm currently working on the idea of Hive Network: Where we can link all previous Hive features to a network for agent, so they can trace entire project history -> the next Hive feature would be improve, more auditbility for human user
And this would align pretty much with your idea on a "remember" agent, instead we gonna introduce new agent for this, we can prefill all "important knowledge" "important memory" to current Hygenic Reviewer with that Reviewer could be more robust and catch more misalignment, halluciation from Planner and Worker. This feature can be toggle by a config as well in case some user don't want this overcomplex stuff.

This should work in theory because we are not making agent "agentic remember" stuff or overload the context window we cherry pick the important part agent need to remember for agent to self improvement which should be hive plan and previous hive context.
Further on implementation if our Hive Network is too big, instead trying to fill every past hive into agent or all document suite, I will introduce a deterministic toolsets for Hygenic Reviewer to navigate Hive Network.

TDLR: I will introduce a knowledge base called Hive Network for all hive and important context in project, the one in having access to that Hive Network for now should be Reviewer, Reviewer will have new skill and toolsets to navigate Hive Network

[Kurounin]

Hi, @tctinh the Hive Network sounds like a good idea. I was recently looking at https://github.com/Fission-AI/OpenSpec/ and their changes history is quite similar to what you're proposing, allowing the agent to supplement the research agent with additional information from older changes that are relevant to the current proposal.

[flatpackfan]

as the "AGENTS.md" serve as psudo-memory it should evolve as project continue would be great way to do memory at the moment.

I think this aligns with my line of thinking about this. In addition to AGENTS.md, my thought was whether it would be useful to extract repeatable patterns into SKILL.md files (that evolve over time). But I'll admit that I'm not very familiar with the limitations of this approach so it might not be a good idea at all.

The thing that I'm seeing is that agents who have to do repeatable work across different sessions struggle to generate consistent output. Capturing those project specific, repeatable patterns somewhere feels like a good idea.

[suse-coder]

Mabye for learning and memory here are also some interesting concepts: https://www.swarmtools.ai/docs/concepts/learning-systems