Negative Knowledge as Failure-aware Shared Memory for AutoResearch

Researchers propose a 'negative knowledge' memory system for AI-assisted research that captures and structures failed experiments as reusable knowledge assets. The approach outperforms baseline AutoResearch systems while reducing token usage, and demonstrates transfer learning capabilities across different scientific problems in nonlinear PDE research.

This research addresses a fundamental inefficiency in AI-driven scientific discovery: failed experiments typically vanish from memory rather than becoming organizational knowledge. The proposed negative knowledge layer operates through two agents—a curator that converts failures into typed, bounded records, and a downstream researcher that explicitly evaluates these records before proposing new experiments. This structured approach differs from conventional error logging by treating failures as epistemic assets rather than debugging artifacts.

The work builds on growing recognition that AI research systems need better mechanisms for cumulative learning. Traditional AutoResearch pipelines often repeat similar failures across iterations, wasting computational resources and tokens. By formalizing negative knowledge as a persistent, queryable resource, the system enables agents to learn from collective failure patterns rather than individual trial-and-error cycles.

The experimental results demonstrate meaningful improvements: the negative knowledge layer solves PDE problems that all baseline approaches fail on, while consuming fewer tokens in same-task retry scenarios. More significantly, the knowledge bank transfers across different PDE problems, suggesting that failure patterns generalize beyond specific research contexts.

For the broader AI research infrastructure, this work implies that scientific discovery systems should explicitly separate and maintain negative findings alongside positive results. This mirrors how human research communities maintain negative registries and meta-analyses. As AI systems become more central to scientific workflows, structured institutional memory around failures becomes critical infrastructure rather than optional optimization. The framework could influence how future research platforms architect their knowledge management systems.

• →Negative knowledge memory converts failed AI research attempts into structured, reusable records that downstream agents can explicitly evaluate
• →The system outperforms vanilla AutoResearch baselines while using fewer computational tokens in both retry and novel task scenarios
• →Negative knowledge transfers across different nonlinear PDE problems, demonstrating generalization of failure-pattern learning
• →Structured negative knowledge should be maintained as collective infrastructure in AI research systems, not merely as debugging aids
• →The approach reduces redundant exploration by making failure patterns explicit and queryable across multiple research agents