Select the parts of a long agent reasoning trace that are safe to prune from context between iterations

An agent's context window between iterations is not a logbook. It is the working surface for the next decision. Everything in it should earn its keep. Four categories are safe to drop. **Consumed tool results** whose conclusions are already in the plan no longer need their raw form; the plan carries the conclusion forward. **Abandoned exploration branches** are dead weight by definition; the agent already decided not to pursue them. **Stale intermediate reasoning tokens** from steps back are the trail of how the plan was reached; the plan is what matters next, not the trail. **Resolved tool errors** from a branch that has since been retried successfully are noise; the success is what the agent will build on. Two categories must stay. The **current plan and next sub-goal** are the agent's anchor for what to do this iteration. The **user's original request** is the goal across all iterations. Drop the goal and the agent forgets the task; drop the plan and it loses its place in the work. The rule of thumb is: keep what the next iteration will read, drop what the previous iteration produced as a means to that end.

**The principle is that an agent's context is a working surface, not a journal.** Anything that has done its work and left a conclusion behind can leave; anything the next iteration needs to read must stay. The four safe to prune categories share a pattern: each one is the means by which a conclusion was reached, and the conclusion is already captured elsewhere. Consumed tool results live on as plan items. Dead-end branches were rejected. Stale reasoning tokens were the cognition that produced the current plan. Resolved errors are superseded by the retry's success. Carrying them forward inflates input tokens, dilutes salience around the live plan, and pushes load-bearing content into the trough of the `lost-in-the-middle` curve. The two keep categories are different in kind. The user's original request is the **anchoring goal** across the entire trajectory. Without it the agent forgets why it is working. The current plan and next sub-goal are the **handoff** from this iteration to the next; without them the agent loses its place. The pragmatic implementation is a structured agent state with explicit slots,`goal`, `plan`, `next_step`, `recent_observations` (last 1-2), `archive` (everything else, queryable but not auto-included). LangGraph's typed state and OpenAI's reasoning-summary feature both reflect this pattern in production.