AGENTS.md was step one.
Coding agents need a written workspace they consult and update, plus a rule for retiring what is stale. Consult, update, retire.
Kevin Liao's fresh take on Hacker News puts it sharper. Coding agents do not need another memory plugin. They need a written workspace they can consult, build against, and update. Operator Memory is one open source illustration of that pattern: a Markdown brain instead of a Retrieval-Augmented Generation (RAG) lottery that surfaces a tiny top slice and quietly forgets the rest.
The loud objection in the thread is real. Docs go stale. Agents append forever. Decisions never retire. Context turns into sludge.
That is not an argument for more recall. It is the same disease Liao names in memory stores. If you do not own the written procedure, including an explicit rule for retiring what is obsolete, you get pollution either way.
What keeps a multi-agent setup sane is consult, update, retire. My own fleet already runs that way: skills, standing procedures, an org chart, living Docs updated in place. Not transcript memory. AGENTS.md opened the door. The operating system is the maintained procedure set.
If your agents cannot find why a decision exists, or cannot retire the file when truth changed, you do not have memory. You have a landfill.