The Pulse
Research Agents Gained 10% by Forgetting Their Search History
A team of four language-model agents improved results by an average of 10% over four parallel agents across four automated research tasks. The study found that periodically clearing agent sessions and candidate histories while keeping the b
AI.info Team · Written with AI ·
Four agents gained 10% by starting fresh
A team of language-model agents improved results by an average of 10% over four parallel agents across four automated research tasks, according to a paper submitted to arXiv on October 14. The team periodically erased its conversational sessions and record of candidate solutions, then restarted from the best result found so far. The approach also beat a single agent by an average of 22% across the tasks, the authors report.
The study, “Scaling Discovery Requires Forgetting,” examines coding and optimization loops in which an agent proposes a solution, receives a score from an evaluator and tries again. The authors argue that as these loops accumulate history, agents grow more conservative: they make smaller edits and progress slows, even when their compute budget is not spent.
“What an agent is allowed to remember matters as much as how much compute it is given.”
Shangyin Tan et al., “Scaling Discovery Requires Forgetting”
The reset keeps the best candidate, not the trail
Forgetting does not mean throwing away all progress. At each reset, the system saves the team’s highest-scoring candidate, but clears the agents’ conversations and the wider collection of candidates, scores and feedback. Fresh sessions begin from that saved solution. In teams of four, agents can inspect one another’s candidates and scores during each round, but do not communicate directly inside their sessions.
The researchers tested the method on AtCoder Heuristic Contest 014, a Frontier-CS optimization problem, a kernel challenge using a simulated VLIW machine, and a small language-model training task. They used the Pi agent with DeepSeek V4.1 Flash, gave each configuration the same total token cost, and ran each configuration three times per task. The team forgot three times, dividing its work into four rounds.
Scores improved across four different tasks
The team with resets produced the strongest mean result on all four tasks. On AtCoder’s AHC014, its mean best score was 1.103 million, compared with 964,000 for four parallel agents and 795,000 for one agent. On the VLIW kernel task, where fewer cycles indicate better performance, the team averaged 1,097 cycles against 1,220 for the parallel agents.
On Frontier-CS Q171, the team scored 80.8 against 70.2 for parallel agents. On the language-model training task, it recorded 0.961 validation bits per byte, compared with 0.980 for both the parallel and single-agent setups; lower is better on that measure. The paper reports an average 10% improvement over parallel agents and 22% over a single agent across the four tasks, while its abstract gives gains of up to 15% and 39%, respectively.
A same-start test points to the reset
To test whether the gains came simply from giving agents more attempts, the authors compared continuation and forgetting from the same plateaued AHC014 state. Both conditions received 30 more proposals per agent. For the four-agent team, forgetting improved the score by 9.2% on average, with a standard deviation of 6.9%, while continuing the existing sessions improved it by 1.4%, with a standard deviation of 1.3%.
The reset condition cost $15.32 per replicate, compared with $24.67 for continued sessions. The paper also tracks the agents’ behavior: before the first reset, the team’s median edits shrank from about 150 lines to 30, and its reasoning-token use fell from about 28,000 to 7,000. Those measures rose after a reset, but they do not establish whether clearing the conversation, candidate history or both caused the gains.
The results leave memory rules unresolved
The findings are limited to four tasks, three runs per configuration and one agent-and-model setup. They do not show that periodic resets will help every research agent or every search problem, and the authors do not compare their method head-to-head with a related approach that starts fresh while retaining the wider workspace. Their own next questions are practical ones: when should an agent forget, and what information should it keep?