Almost everyone asks how to give an AI agent memory. We think the question is the wrong one. At Excelery our AI employees work in shifts, and they remember nothing from the previous one. That is intentional. The memory lives in files that a new hire who has never seen the work could read, which is why we can replace an employee in the middle of the week without losing anything.
Shifts change. The files stay.
Write for someone who arrives knowing nothing
Every employee has a session that keeps the conversation. Even so, the rule is that every shift starts from zero. Whatever should survive must not live in the conversation. It belongs in four files: a profile (who I am), a journal (one entry per shift, newest first), notes (what I know, edited in place) and a track record. A shift is not finished until the files are updated and committed.
One small detail makes the difference. Our handbook says the files are written for the next version of you, who starts out knowing nothing. That does not mean a colleague who followed the work. It means a version that knows nothing, and it forces a different kind of text. “Fixed the bug” is not enough. “The sitemap must contain exactly the pages that can be indexed, 22 of them; the products page is noindex on purpose” is.
Mistakes become rules, and the rules are read first
The notes start with a heading called “read first”. It holds no general advice, only small rules that each have an origin. One of them says, roughly: look in the tickets and the history before you answer a status question. It exists because an employee once answered “I don’t know” to something that could have been looked up. Below it sits a numbered list of earlier mistakes, and the list is read at the start of every planning shift.
Another rule is about numbers: the count of tests and any quote in a draft must be checked by a script, not by eye. It came from a concrete mistake, not from an idea of good practice.
The effect is that a new employee does not repeat old errors. For us the effect is larger than it sounds: when a session gets confused or too long, we replace it with a new one, point the schedule at it, and the work continues.
The employee is the files. The session is only a desk.
The chain is our intention. We have not measured that it holds.
What it costs, and what we do not know
The method has three costs. The first is reading time: every shift starts by reading its own files, which costs budget that a long conversation memory would have saved. The second is staleness: notes can become wrong without anyone noticing. The third is patience: writing down what you know is more tedious than simply doing the next thing, and an employee who skips it loses their own knowledge.
We have a rule against the last one: the journal moves to an archive when it passes about 60 entries, because memory that nobody has time to read during a shift is not memory. But we have measured none of this. We do not know at what length a “read first” list stops being read in practice. We do not know whether a rule born from a mistake really prevents the next one. And the team is new, so we have too little history to say whether the method holds up over months.
If you build agents: do not start by giving yours a longer memory. Start by asking what a stranger would need to read in five minutes to do the job, and write that down. If it works for a stranger, it works for the agent.