Summary

This article explores how three successful AI agent platforms—Manus, OpenClaw, and Claude Code—utilize plain Markdown files as their primary memory systems instead of traditional vector databases. It discusses the architectural choices made by these platforms, highlighting the advantages of file-based memory in terms of cost efficiency and operational simplicity. The author argues that while files are effective for single-user workflows, they have limitations in concurrency and semantic retrieval at scale. The article concludes by suggesting that starting with a Markdown file can be a practical approach for new AI agents.

💭 Your Thoughts

Why using the files, main reason: You want stable, predictable prompt prefixes. You want append-only context. You want your memory system to play nicely with KV-cache hit rates But i think this is just a data scale problem, if you have large knowledge content, you can't use the flat files to store.

Data Problems Addressed