Agents don't need RAG memory plugins — they need Markdown docs they can update

Kevin Liao's argument is blunt: the whole "agent memory" product category is solving the wrong problem. Most plugins scrape transcripts, carve them into a thousand vector snippets, and hope similarity search picks the five that matter on the next prompt. That is still just RAG with extra dreamers and rerankers burning tokens overnight.

https://liao.gg/blog/agents-dont-need-memory

https://news.ycombinator.com/item?id=49945933

His list of failure modes is familiar if you have tried any of them. Similarity is not correctness. Snippets lose motivation and environment. Yesterday's auth note stays "true" after the code changed. Agents cannot search for what they do not know they forgot. And a opaque SQLite of embeddings is basically unauditable when something goes wrong.

The alternative he has been running for a year is boring on purpose: a Markdown brain the agent consults before work and updates after — specs, decisions, indexes, research — committed like any other project file. No embeddings. No overnight curators. Readable by humans on the team.

HN split between people who called this AGENTS.md with discipline and people who still want automatic recall so they never have to curate. Fair fight.

If your coding agent keeps reinventing decisions you already made, would you rather bolt on another memory plugin or force it to maintain docs like a junior who writes things down?