A November seventh preprint, Think or Remember, examines a related serving interaction: deleting old reasoning can break an exact-prefix cache. The authors forked a hundred and three matched Qwen Three eight-billion-parameter agent states, then let Keep and Drop versions continue independently. Under modeled generated-state reuse, Drop cost one point seven four times Keep in operations. Under no reuse, the ratio fell to zero point three four.

Those continuations also generated different behavior, so that comparison mixes cache reconstruction with the model's response to losing its history. A separate mechanical replay held continuations fixed. Keeping only the latest five reasoning blocks cost about four point five four times retaining everything, despite nearly equal prompt volume, under its specified unlimited exact-prefix cache. Removing an early block forced later material to be processed again.

These are modeled operation costs, not elapsed times or API bills. Quality also depended on the setting: dropping reasoning lost nine to twenty-three percentage points on live HotpotQA runs, while the gpt-oss ScienceWorld result favored dropping. A history policy changes both the reusable computation and what the agent subsequently does.

Source 1