Self evolving CUGA?

[merveunu]

It would be great to see CUGA learn from the past -- whether from failures or from common knowledge it acquires over time. What are the plans to make CUGA self-learning as we run it

[aviyaeli]

@merveunu this is a great suggestions.

We are experimenting and developing technologies in the area of trajectory analysis and summarization so CUGA can learn from experience. The idea is that these experiences get stored into Long Term Memory so the agent can adapt with time.

There's tension between the ability to learn and adapt, and the expectation for reliable deterministic behavior. We believe that policies and human-in-the-loop mechanism could help balance this tension in a trustworthy manner.

@visahak fyi

[visahak]

100%. CUGA has a feature where it tries to generate insights from the agent trajectory and stores it in LTM to be used as guidelines for similar tasks. Would love to discuss more on how to enhance it and augment it with additional requirements.