How I run my own work with agents
What happens when the AI agent forgets what we agreed?
The short answer
It carries on without the agreement, and it does not tell you. An agent remembers only what is written somewhere it reads before it starts. Anything agreed only inside a conversation is lost when that conversation ends or grows too long. The fix is to write decisions down in a place the agent must read, not to hope it remembers.
An agent starts every conversation new
A colleague who agreed something with you last week still knows it today. An agent usually does not. Each new conversation starts empty, except for what you give it. Even inside one long conversation, an agent can hold only so much at once. In many products, as the conversation grows, the oldest parts are shortened or dropped.
The danger is the same as with confident wrong answers: the agent does not say “I have forgotten.” It works on, fluently, without the agreement.
What I learned in my own system
My follow-up agents run every few hours, and every run starts fresh. From the first day, my standing worry was that both I and the agents would forget what had been agreed.
So the system keeps a written record: every decision with its date, every open item, every draft and what happened to it. Each run begins by reading that record. Nothing depends on an agent remembering. And once a day, a separate agent checks that the agents’ instructions still match the decisions in the record.
More memory is not the answer
The tempting fix is to give the agent everything: every email, every past conversation. That makes things worse. The more you load into an agent at once, the more it misses inside what you loaded. In my own system, I treat that load as the biggest single risk to accuracy.
Good memory for an agent is short, current and organised. It holds decisions, not transcripts.
What this means for you
- Every agreement that must last goes in writing, in a place the agent reads before it starts work.
- Keep that record short. When a decision changes, replace the old one. Do not pile the new one on top.
- Check that the agent actually read it. Early in the work, ask it what it knows about the decision that matters most.
An agent’s memory is a document that someone maintains. If nobody maintains it, the agent has no memory.