That's what I was thinking about when I lay in bed one more time while trying to fall asleep.
I was sick of my agents forgetting stuff that I told them a few days ago due to compressing their memories. With Gemini it was even worse, sometimes the context just got reset and it wouldn't remember things from 5 minutes ago.
So I thought about how I organize my memories. We don't have all the things we've experienced in mind at the same time. We keep them in junks. When you think about "holidays" you will remember all the holidays you've been to. Then you can dig deeper into one memory and think about that trip to Rome. Go even deeper and you'll remember the hotel, if the bed was comfortable, how the food was, etc. Also you'll remember the people who were there with you and when you think about those, there will be many more memories that connect with them. Things you hadn't thought about 3 minutes ago, but they were there in your memory. Just on demand.
Then I realized, this concept should be transformable to an MCP. Next morning, I told Claude about this idea and we made a prototype. Well, and it pretty much nailed it. We created skills to properly use the MCP and what should I say... I sometimes see it using it's hmem-MCP without me asking to do it. It reads and writes memories on its own.
Imagine that. No more inefficient .md-memory-files that will flood the context and waste your tokens.
Check my Github if you want to know how it works.
It's still in development but it already works pretty well! If you pass the Github-link to your Openclaw, it will be able to install and use it.