On October 3, Kevin Liao published an essay titled "Agents Don't Need Memory. They Need Documentation." It reached the Hacker News front page and drew 297 points and 173 comments, according to AgentConn's analysis of the debate. The essay was about coding agents. Its argument reaches further. It is a direct challenge to how institutional knowledge AI is usually built: capture everything, store fragments, and hope retrieval finds the right one later.
The argument, stated fairly
Liao's case, in its strongest form, is simple. Most memory add-ons do the same thing. They read session transcripts, cut them into short fragments, store the fragments in a vector database, and attach the few most similar fragments to each new prompt. If the agent is still confused, it gets a search tool to look for more.
That design has five faults, as he sets them out in the original essay. Retrieval ranks by similarity, and similarity is not correctness. A fragment loses the setting it came from: why a choice was made, what was rejected, and what lesson followed. Old fragments stay in the store after the facts change, so the past is treated as current truth. An agent cannot search for something it does not know exists. And a store of thousands of embeddings cannot be audited. No one can say which entries are stale, which were never used, and which are wrong.
His alternative is a written record the agent reads before work and updates after it: instructions, specs, research notes, and an index that points to the right document. People work this way already. No team replays a three year old meeting to recover a constraint. The team reads the note that was written at the time.
Why institutional knowledge AI should listen
Companies have the same fault at a larger scale. The record of a decision sits in a chat thread, a call transcript, a ticket, and a slide deck. A search tool returns fragments from all four. The fragments may be similar to the question. That does not make them current, approved, or complete.
The documentation route has evidence behind it in software. AGENTS.md, a plain Markdown file that tells coding agents how a project works, has been adopted by more than 60,000 open source projects, a figure OpenAI reported when the format moved to the Agentic AI Foundation. One file, written on purpose, read by every major coding tool, versioned with the code. No database and no retrieval step. The format spread because the record is inspectable. A person can open it, read it, and fix it.
That inspectability is the point companies miss. A knowledge base that employees can open and read is a different thing from an index only an AI can query. The first can be checked in a meeting. The second has to be trusted.
The pushback, and what it proves
The Hacker News thread did not accept the essay whole. A detailed review of the essay and the discussion on explainx records the main objections. Agents write poor documentation when left alone. They record guesses as facts, keep one off comments as if they were general rules, and let files grow until no one reads them. Documentation also goes stale if nobody owns it. And no one in the thread produced a head to head test of a documentation store against a vector memory store on the same tasks. The claim is argued, not measured.
Those objections do not rescue fragment retrieval. They set the condition for the documentation route to work: review. A record earns trust when a person with the right to decide has read it and approved it, and when a later change replaces the old version instead of sitting next to it. That is the model set out in How Jilu Works, where captured material passes a draft, review, and publish gate before it counts as company knowledge, and where a new fact can extend an old record without erasing it.
Capture is the staging area, not the record
The useful split is this. Capture everything, because decisions are made in chat, calls, and documents, and no one will stop to write them up. Then promote a small part of it, through review, into a record that names what was decided, why, by whom, and what it replaced. Memory systems stop at the first step. Documentation, as Liao argues for code, is the second step done on purpose.
Our earlier post, The AI Agent Memory Race Is On. Companies Are Behind., made the case that companies lag agents on memory. This debate sharpens it. The race that matters is not who stores the most fragments. It is who turns captured work into records that stay current, carry permissions, and can be read by the people they describe.
That is the standard institutional knowledge AI should be judged by. Not how much it remembers. How much of what it remembers a person can open, check, and stand behind.