Scaling Discovery Requires Forgetting
Autoresearch LLM agents plateau when their growing session history makes them conservative. This paper shows that periodically discarding everything except the best candidate—forgetting—resets exploration and unlocks further gains. The finding suggests that how an agent manages memory is as important as how much compute it receives.
The paper makes a strong case for a minimalist intervention. Its four-task evaluation and mechanistic tracing of edit sizes and reasoning tokens show that accumulated context causes plateauing and that periodic forgetting unblocks progress. However, the work is limited to a single model and harness, the forgetting schedule is hand-tuned rather than optimized, and no code is released.
The core insight is both simple and well substantiated: agents narrow their edits and slash reasoning as sessions lengthen, and forgetting resets both behaviors. The cost analysis is a nice bonus—post-forget sessions are cheaper because shorter context reduces token spend. The within-run ablation continuing from the same plateaued state (Table 3) is a particularly clean control.
The forgetting schedule is arbitrary: three fixed forgets are used for every task, yet Table 2 shows that most improvement comes from the first and returns diminish sharply. Whether adaptively timed forgetting would help or hurt is unexplored.
The experiments rely on a single model (DeepSeek V4.1 Flash) and a single harness (Pi), and all tasks are code-optimization problems; generalization to other agent frameworks or scientific domains is unproven. The paper also does not empirically compare against simpler diversity baselines such as temperature scaling or independent random restarts, leaving open whether the gains are specific to forgetting or just to fresh-session variance.
The main evidence is persuasive but not exhaustive. Comparisons between the single agent, parallel agents, and team with forgetting are conducted at matched total token cost, which is fair. The continue-versus-forget ablation from an identical plateaued start (Figure 4, Table 3) is the strongest evidence that the intervention itself drives gains. However, only three replicates per task are reported, and the single-agent baseline shows large variance, so the reliability of the measured improvements is unclear. The distinction from concurrent work is conceptually sharp but not empirically tested: the authors accurately note that Fork-and-Flush inherits the entire workspace, whereas they keep only the best candidate, yet this difference is not evaluated head-to-head.
Reproducibility is a notable weakness. The authors describe the harness and prompts (Appendix B), but they do not release the implementation, the reducer logic, or the peer-communication code. Hyperparameters such as sampling temperature and top-p are not reported. Only three replicates per configuration are run, which is low given the high variance of LLM agents, and budgets are reported in dollars rather than tokens, tying results to a specific API pricing regime. Independent reproduction would require rebuilding the entire harness from the paper description.
Automated research with LLM agents has shown promising results in various domains. Through iterative refinement and scored verifiers, LLM agents are capable of discovering new ideas. There are different ways to scale automated research: let one agent iterate for longer, run several agents in parallel, or further let the parallel agents communicate and build on one another's candidates. However, we find that all three strategies eventually plateau as the agents accumulate history. Each agent becomes anchored to its own past attempts, making fewer and smaller edits. Beyond this point, additional budget brings little further improvement. In this paper, we show that the key to scaling discovery further is to forget. Periodically discarding the search state, including every agent's session and the candidate history, and restarting from the best candidate found so far, unblocks the search and pushes the same budget further. We implement forgetting in a team of agents with communication and evaluate it on four tasks spanning heuristic programming, kernel optimization and model training. The results show up to 15% improvement compared to parallel agents, and up to 39% compared to a single agent. Our results suggest that, for open-ended discovery, what an agent is allowed to remember matters as much as how much compute it is given.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.