Should an AI agent keep a long-running context, or is it better to restart tasks from a clean state…

I would use a hybrid: clean context per subtask, persistent state across subtasks.

A long-running agent should retain the goal, accepted decisions, constraints, links to artifacts, and unresolved questions. It should not retain every intermediate thought, failed attempt, or raw tool log in the active prompt. Those belong in external files or a task database and can be retrieved when needed.

This is supported by two observations. The “Lost in the Middle” research shows that models do not use all positions in a long context equally well. Anthropic’s work on long-running agent harnesses describes compaction and explicit progress handoffs as ways to continue across context windows.

I would restart or compact when the task changes, the working set becomes noisy, or the agent starts repeating itself. I would continue the same run when the next step depends on a small set of verified decisions and artifacts.

The key is not “memory versus reset.” It is durable state versus transient context. A good checkpoint should let a fresh agent resume safely without pretending that the entire transcript is equally important.