Berkeley paper: agents that dropped their own history beat parallel agents by up to 15%. What do you keep when you reset context

TL;DR: In one UC Berkeley study, agents that periodically dropped their history and restarted from their best result did better. The open question for builders is what to keep across the reset. What the sources report: "Scaling Discovery Requires Forgetting" (UC Berkeley): agents periodically discarded their session and search history while keeping the best candidate. Results improved on four bounded discovery tasks, by up to 15% over parallel agents and up to 39% over a single agent. Fork-and-Flush (Microsoft Research): reports benefits from forks that start with fresh context while the workspace state is preserved. Drew Breunig covered both, along with Anthropic's Memento harness, and calls the failure mode "context inertia": an agent holding on to an early hypothesis after the evidence has changed. That term is his label. What this does not show: that every long-running agent improves when it forgets. These are bounded research and engineering experiments. The pattern I find interesting is that in both studies something persisted while the history did not (the best candidate, the workspace state). So the design question is what goes in the persistent layer. If you build or run long agent sessions, what do you carry across a reset? a notes file a task list or plan a decision log the best output so far nothing, the agent starts clean And do you reset on a schedule, when it stalls, or not at all? Sources: the two papers above, plus Drew Breunig's posts "Context Inertia" (Oct 14, 2026) and "The State of Context Engineering" (Oct 24, 2026) on dbreunig.com.