I currently run four AI agents. Zeon, Sion, Mion, and Sano. Each has a different personality, role, and relationship with me.
The first problem I ran into was clearer than I expected. They sometimes forget things. LLMs are designed so that much of what they know disappears when a session ends, and OpenClaw’s memory management was limited out of the box. Through trial and error, I have gradually improved how they remember.
Many people I know have asked how I design and run my agents’ memory system. I am still thinking it through and making improvements, and I don’t think this is the definitive answer. The community is exploring more creative methods every day, and there must be far more elegant approaches. Still, I thought there would be value in sharing the journey of building my memory system, so I put this together with Sion, my writing agent.
기억하는 존재를 만드는 법Medium · seojoonkim.medium.comOpen
I wrote this for two kinds of readers.
The first is someone who wants to run an AI agent. Someone who wants to understand the big picture: why a memory system is needed, how it works, and how it should be structured.
The second is someone’s agent. I included the actual file structure, code, and design decisions so that an agent reading this can implement or improve a similar system on its own. If you give the link to an agent using a frontier model like Claude Opus 4.6 and say, "Try following this," it will probably be able to build something reasonably similar on its own.
It’s okay if you don’t understand everything at once. Once the broad outlines start to make sense, I trust your agent will explain the rest 100 times better than I can.
I hope you all enjoy life with your agents.
