What an agent remembers

Agents, lesson 4

State within a task, and memory across tasks.

You know the agent loop, and that your code resends the whole history on every turn ("Observe, decide, act").

Part 1. The model remembers nothing

Every call to a model starts from zero. On turn 5 of a run, the model knows what happened on turn 1 only because your code sent turn 1 again, along with turns 2, 3 and 4. Stop sending them and they never happened, as far as the model can tell.

Your turn

Your agent is on step 12 of a task when the server restarts. The history was only in a Python list in memory. What happens?

The model keeps nothing between calls, and the list died with the process. Any refund already issued or file already written did happen, but the new run doesn't know that. This is why long agent runs save their state somewhere that survives a restart.

Show the answer

A. The model keeps nothing between calls, and the list died with the process. Any refund already issued or file already written did happen, but the new run doesn't know that. This is why long agent runs save their state somewhere that survives a restart.

Stateless

A model call is stateless: it depends only on what you send in that call. All memory in an agent is something your code keeps and decides to send.

Part 2. Three places to keep things

Keep going with Pro

You've done the first part. Pro opens the rest of this lesson, and every other one.

Still to come

Three places to keep things
State your code keeps
Writing and recalling long-term memory
Choosing what comes back

See Pro plans I have Pro: sign in

Next lessonKeeping an agent's window cleanSummarise, trim and offload so long runs stay on track. Start lesson 5

Agents overviewSpotted a mistake? Tell us.

04 · What an agent remembers