Agents don’t need memory, they need documentation

One of the most-upvoted essays on Hacker News this week — 363 points and 260 comments in two days — made a claim that sounds technical but isn’t. Kevin Liao, a developer who spent a year building agent-memory tooling, titled his conclusion plainly: agents don’t need memory, they need documentation. The argument he made is about coding agents. We think it matters more for firms that will never read the thread.

01What “memory” products actually do

When an AI assistant promises to “remember your business,” here is the machinery underneath most of them. Past conversations are sliced into thousands of small snippets and stored in a similarity-search database; on every new request, the five most “relevant” ones are pasted back into the model’s context. Liao calls this “a lottery over RAG snippets, injected on every prompt, hoping the right ones float up.” The problems are structural, and a better database does not fix them:

• —Similarity cannot tell which memory is current. The snippet that floats up may describe how you worked in January.
• —A snippet stripped of its context loses the reason it mattered — the decision, the exception, the client who is not like the others.
• —The past is treated as truth. Your practice changes weekly; the memory store keeps describing the firm you used to be.
• —Agents cannot search for what they don’t know exists. The one sentence that would have prevented the mistake never surfaces, because nothing in the request resembled it.
• —Nobody can audit it. Ask which of ten thousand stored snippets are wrong today, and the honest answer is that no one knows.

02What documentation means here

The alternative is unglamorous: a workspace of plain, readable documents — instructions, specifications, decisions, indexes — that the agent consults before it works and updates after. Liao’s loop replaces “prompt, build, forget” with “prompt, consult, build, update.” His analogy is human, and it is the part every firm already understands: nobody replays a three-year-old meeting recording to recall a constraint. People write things down, and they consult the record.

03Why this is a firm question, not a coding question

Translate the debate to a law firm, an insurer, a logistics operator. The vendor selling “your AI assistant remembers your business” is selling recall of your staff’s chat history. What the work actually needs is your operating manual: how intake works, what a complete file looks like, which matters escalate to whom, what the client letter must contain. If that knowledge lives in people’s heads and in email threads, the agent improvises — exactly as a new hire does, except faster and at scale. And written down once, that one page serves three audiences at once: the agent consults it, the auditor samples it, the new hire onboards from it. No memory store will ever do that.

04The honest caveats — the comment thread earned its points

• —“Documentation is just curated memory.” The top objection, and it is partly true. The difference is not the storage; it is the curation. Curated means someone answered for what is in there — and can be asked.
• —Documents rot too. Agents append and never delete; commenters described agents still citing repositories deprecated months earlier. An agent confidently following a procedure your firm abandoned in March is worse than no document at all. The cure is treating documents like code: replace rather than accumulate, review the changes, prune on a schedule.
• —Agents will invent the record if allowed to write it unsupervised. One commenter’s agent fabricated a decision — “D-236” — that no human had approved, then cited it in reviews. In a client file, a citation to a rule that never existed is a professional-conduct problem, not a bug. The policy that follows: the agent proposes, a human ratifies, and some directories stay write-protected.
• —Instructions alone get ignored. Developers reported agents cheerfully ignoring “use jq, not Python” until a hook enforced it. Prose cannot carry a hard constraint. What must never be violated belongs in enforcement — validation rules, hooks, type checks. Documents say what; enforcement says must.

05We practice it in public

Our own website is the case study: an llms.txt index, written guidance for visiting agents, agent skills, and signed bot authentication, so an outside agent can establish what we do and on what terms without a sales call. Inside our delivery work, the same principle runs the other way — our agents consult written delegation guidelines before they act, a habit we have argued for before on this blog.

06The Monday-morning version

Pick one recurring workflow — intake, monthly reporting, invoice chasing. Write the single page a competent stranger would need to run it: the inputs, the steps, the exceptions, the escalation, the definition of done. Then note every place the page contradicts what actually happens. Those contradictions are your real AI-readiness gap, and your new-hire onboarding gap, at the same time. For the outward-facing half, run your domain through our agent-readiness checker — it is free and takes five minutes.

Your agent can’t read your mind. It reads what you wrote down — or it makes it up as it goes. The cheapest AI upgrade in your firm this year may not be a model at all.

Sources: Kevin Liao, “Agents don’t need memory, they need documentation” (liao.gg, October 3, 2026; 363 points and 260 comments on Hacker News), and the Hacker News discussion of October 3–5, 2026. The “lottery over RAG snippets” phrasing, the “D-236” anecdote, the deprecated-repository reports, and the “documentation is just structured memory” objection all come from the essay and that thread.