Failure Broadens, Success Sharpens: Continual Adaptation of Tool-Use Agents Across Diverse Tasks with Dual-Polarity Memory

Large language model agents can self-improve over time by utilizing procedural memory derived from prior task trajectories, yet the distinct utilities of memories derived from local successes and failures remain underexplored. This work studies this question in a realistic setting with heterogeneous streams of multi-domain tool-use tasks and various environment constraints common in real-world scenarios. We develop $\textit{Dual-Polarity Memory with Periodic Maintenance}$ (DualPM), a simple, retrieval-free memory workflow characterized by (1) customizable reflection and procedural memory generation from both local successes and failures within each trajectory, and (2) periodic memory consolidation over long task streams under a given memory capacity, preventing memory explosion. Online prequential evaluation with fine-grained metrics reveals that failure memory more strongly broadens coverage, success memory better streamlines execution, while combining both polarities effectively integrates both benefits. Further studies demonstrate that DualPM enables cross-domain memory generalization and continual memory adaptation, while external evaluation on EvoMemBench reveals the portability and effectiveness of DualPM as a general long-term procedural memory method.