Fresh-Context State Summary

Rebuild each agent turn from a written environment summary so the only memory of prior steps is what the harness deliberately records.

Some browsing and tool agents reset the model prompt every step: the LM sees a fresh context containing the question, current page (or tool state) at the cursor, and a harness-written summary of prior actions — not the raw full trajectory. Prior steps persist only if the summary records them.

This is deliberate context engineering and agent memory tiers: the summary is the durable working memory; observations stay local to the turn (pair with collapsed observations when you keep a longer transcript instead). Design the summary schema carefully — missing fields are permanent amnesia; the same fidelity trade-off governs whole-session resets (compaction vs context reset). Combine with a small discrete browser tool action inventory so each turn still emits one valid command under action format enforcement.

---

Sources

• WebGPT (Nakano et al.), §2