Negative Knowledge as Failure-aware Shared Memory for AutoResearch

AI-assisted research systems generate many failed attempts, but those failures rarely become durable shared knowledge asset. We propose a negative knowledge memory layer: a curator agent converts each failed attempt into a bounded, typed record in a shared bank, and a downstream research agent must explicitly adopt or reject those records before proposing its next experiment. We evaluate the layer in two settings: same-task retry and cross-task scientific research. On the deterministic subset of ScienceAgentBench, the negative knowledge layer outperforms baselines while using fewer tokens. More importantly, in two nonlinear math physics problems with coupled PDE systems, the bank enables cross-task reuse of past failure boundaries: agents with the bank solve new tasks that all no-bank baselines fail, with partial transfer to a different coupled PDE system. These results suggest that structured negative knowledge is not merely a memory-compression or debugging aid; it is a knowledge asset that should be explicitly maintained in broader AI-engaged scientific research, forming, alongside positive findings, a collective infrastructure for scientific memory.