Keeping agents’ knowledge current without hand-maintaining a wiki.
A folder of plain Markdown notes that a person and an agent both write into, and a store underneath that turns those notes into something an agent searches at the start of a session. I didn’t write the memory software; I designed the system around it.
What needed
solving.
I run agents on real client accounts, and most of what they need is institutional knowledge: who decided what, which date is the real one, what was tried last month and didn’t work, what a client won’t accept. An agent that doesn’t know what happened last week is worse than no agent, because it answers anyway. I knew a hand-kept wiki would fall off my priority list within a couple of weeks, so the question became how to keep the notes honest without a person doing it.
What I decided,
and why.
- 01
Procedure goes in skills; state goes in the vault. Skills hold how to do things. The vault, the folder of notes, holds what is true right now.
- 02
Plain text files are the only source of truth. Everything downstream can be rebuilt from them, and nothing writes back into the notes. A personal journal folder in the same vault is excluded from every import, which is the reason I’m willing to point agents at the vault at all.
- 03
Each project gets its own memory, named rather than guessed. Retrieval is scoped before the question is asked. An empty answer is information. It is not an invitation to go looking somewhere else.
What shipped,
and what isn’t proven.
A nightly job loads only new or changed notes and turns them into structured facts, and newer notes supersede older ones instead of deleting them. Twice it failed without saying so: once when a sync path pointed at a shortcut instead of the real folder, and once when agents working in throwaway folders filed lessons under project names that vanished an hour later. Both were caught by a person reading the output, not by the system reporting. Not proven: meaning-based search isn’t pulling its weight yet, and I’d rather say so than claim it.
What I’d do
differently.
Build for replacement earlier. The notes are plain text for that reason, but everything that makes memory arrive automatically is wired into one AI coding tool. And verify that it works, not that it ran: a handful of questions I already know the answers to, asked on a schedule, would have caught both failures.