Berkeley paper: agents that forget their history beat parallel agents by up to 15%
ombre-ai · reddit · 2026-10-28
A UC Berkeley paper, "Scaling Discovery Requires Forgetting," shows agents that periodically discard session and search history while keeping their best candidate improve on four bounded discovery tasks — up to 15% over parallel agents and 39% over a single agent.
- Microsoft Research's Fork-and-Flush reports similar gains: forks start with fresh context while workspace state is preserved.
- Drew Breunig labels the failure mode "context inertia": an agent clinging to an early hypothesis after the evidence has changed.
- Caveat: this doesn't mean every long-running agent improves by forgetting — these are bounded experiments. The real design question is what to persist across a reset (notes file, task list, decision log, best output, or nothing).
The author asks builders what they carry across resets and whether they reset on a schedule, on stalls, or never.
More from coding & agent
- Free Claude Code open-sources multi-provider model swapping for Claude Code, 61 providers — JafarNajafov · 2026-10-28
- Only Then Labs prototypes re-checking agent approvals when underlying evidence changes — tallmetommy · 2026-10-28
- Google releases free 2-hour workshop on multi-agent systems, from first agent to graph orchestration — rseroter · 2026-10-28
- Omarchy 4.5 RC2 ships per-monitor workspaces, floating layout with dock, and 4 new apps — mark_k · 2026-10-28
- Claude Catches North Korean Supply Chain Attack Disguised as an Auth Error — banteg · 2026-10-28
- Expo ships agent skills collection for AI coding tools, already 3.3k GitHub stars — expo · 2026-10-28