The agents can keep the to-do list

Working with agents gets so many things done that people start forgetting about them. By the evening, half of what changed that day has already slipped away. How many times have you thought: “wait, I asked ChatGPT about this some time ago”? We remember that we asked, and rarely what came back.

Bluma Zeigarnik noticed almost a century ago that a finished task lets go of the mind, while an unfinished one keeps pulling at it. The memory part of her 1927 result has been hard to repeat, since unfinished tasks aren’t recalled better overall, though people do tend to go back to them when they get the chance, according to a 2025 meta-analysis from the University of Bern. Agents finish things so fast that most of them never pull at all: we didn’t hold the task long enough for it to stay with us, and since we didn’t do the research, thinking or coding, there was very little to make it stick. Most people who wrote an essay with ChatGPT couldn’t correctly quote a single line of it minutes after finishing, in a 2025 MIT Media Lab preprint.

And the labs are pushing toward an even bigger software factory, with agents running for hours in parallel and shipping thousands of changes. One team shipped about a million lines of code in five months without writing a line by hand, OpenAI reported in February. They’ve noticed the memory problem too, and they’re solving it for the agents: managed agents now “dream” between sessions, rewriting what they remember, as Anthropic announced in May. Dreaming is a scheduled process that reviews an agent’s past sessions and curates what it remembers for next time, as Anthropic describes it.

Maybe that’s the right split. The small things were always meant to fade once they were done, and an agent holds them better than we do. What should stay open in our heads is the goal, the one loop that never fully closes, and checking the work against it every so often is the only way to tell whether a thousand changes nobody remembers still add up to what we wanted. In the same release, Anthropic shipped a feature it calls outcomes: a person writes a rubric, a separate grader checks the agent’s work against it, and the agent takes another pass until the output is good enough.

The agents can keep the to-do list. The goal is ours to keep.