Blog Debate: Agents Need Documentation, Not Memory

Kevin Liao, who writes the blog liao.gg, argues against the current trend of giving AI agents "memory" plugins. These plugins turn your past chats into fragments. The fragments get stored in a vector database, which is a search index that matches text by meaning. Then the system feeds the five most similar fragments into every new prompt. His complaint is this: you cannot check whether a fragment is still true. It never goes through a pull request, which is a tracked, reviewable code change, the way real code does. The result is that different agents end up with different memory, and none of it is verified. His fix is not one short instructions file. He says plainly that a single AGENTS.md file is not enough. Instead he keeps a whole folder of plain markdown documents: specs, decisions, research notes, and an index. The agent reads these before starting a task. Afterward, the agent updates them by hand. There is no automated process that does this for it. He built this into a free open-source plugin called Operator Memory.

I agree with him on the main point. Documentation that lives in the repo gets checked the same way code does. A memory store does not get that check. This matches a problem I have run into myself: an agent that follows a vague instruction can look fine and still be wrong. His habit from the post of writing exact commands, like 'pnpm vitest run <file>' instead of "run the tests," is a concrete example of that same discipline, not a side detail. Here is where I would be cautious, though. Trusting an agent to keep a whole folder of documents up to date on its own, with no automated check for staleness, depends heavily on whether the agent actually does that upkeep every single time.