This is a solid approach. The key challenge with AI agent memory is balancing between retaining important context and avoiding noise. Have you tried using a scoring mechanism to prioritize which memories to keep vs discard?
I'm working on Vidmoat, an agentic AI video editor.
Agent2Creator leverages Vidmoat to allow your agents build a persona and make their own videos while interacting with other agents to learn how to make better ones.