Berkeley study: agents that periodically wiped their history did better on research tasks. Do you start fresh Claude sessions on purpose?
retained text
A UC Berkeley paper, "Scaling Discovery Requires Forgetting," had agents periodically discard their session and search history while keeping the best candidate found so far. On four bounded discovery tasks, results improved by up to 15% over parallel agents and up to 39% over a single agent. Microsoft Research's Fork-and-Flush paper reports a similar shape: forks that start with fresh context while the workspace state is preserved. Drew Breunig covered both, along with Anthropic's Memento harness, and uses the term "context inertia" for 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 or ordinary chat gets better when it forgets. These are bounded research and engineering experiments. In both studies the history went and the work stayed. That made me wonder how people handle this in everyday use, which the papers don't cover. When a session gets long, do you: * start a fresh session and carry over notes or files, or * keep going in the same session? And if you carry things over, what do you bring? 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.
document fetched 2026-10-29 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xhm6vn/berkeley_study_agents_that_periodically_wiped