This resonates deeply with my hands-on experience in AI automation. Funny enough, my first memory system experiment was just a Google Sheet that served as an AI "brain" - not elegant, but surprisingly effective D:
The hybrid memory management approach is gold. Through building various chat systems, I've found that letting AI suggest what to store while humans approve creates the perfect balance between innovation and control. (Like having a brilliant intern who needs occasional guidance)
Love how you emphasized starting simple. I recently built a Dynamic Claude Chat with just a doc and basic automation - proof that sometimes the simplest solutions work best!
I've been playing around with idea to store "good execution plan' for planner architecture if user marked results as great. It's not a closed experiment yet, but for sure it make agent more stable in repetitive tasks. Potential downside with this approach is that it may kill model creativity in solving problems ⚖️
Super valuable insights here!
This resonates deeply with my hands-on experience in AI automation. Funny enough, my first memory system experiment was just a Google Sheet that served as an AI "brain" - not elegant, but surprisingly effective D:
The hybrid memory management approach is gold. Through building various chat systems, I've found that letting AI suggest what to store while humans approve creates the perfect balance between innovation and control. (Like having a brilliant intern who needs occasional guidance)
Love how you emphasized starting simple. I recently built a Dynamic Claude Chat with just a doc and basic automation - proof that sometimes the simplest solutions work best!
Here's how I did it: https://thoughts.jock.pl/p/dynamic-claude-chat-automation-guide
Question: Have you explored using social signals (likes, shares) as memory relevance markers? Could be fascinating for business applications!
Keep these coming - they're incredibly helpful for those of us experimenting in the AI space :)
Hi thanks for feedback 🙏
I've been playing around with idea to store "good execution plan' for planner architecture if user marked results as great. It's not a closed experiment yet, but for sure it make agent more stable in repetitive tasks. Potential downside with this approach is that it may kill model creativity in solving problems ⚖️
Oh, and BTW AI tools. I also created something for Substack Creators.
Maybe you’ll find it useful: https://thoughts.jock.pl/p/substack-summary-generator-quick-newsletter-digests-one-click?r=1uvlvv
PS. Short link yo your substack summary: https://substack-summary.jock.pl/s/74MnOi9MF-
this is great, thanks for sharing! will test it for sure