Title: MemPro: Treating Agentic Memory Systems as Evolvable Code

Abstract: Autonomous agents operating over long horizons depend on robust memory architectures to preserve past data, monitor shifting states, and leverage relevant insights that extend beyond the limits of finite context windows. While current agentic memory solutions generally rely on a memory construction-retrieval (MCR) workflow, they typically focus their adaptation efforts solely on the memory bank, leaving the surrounding pipeline static after initial deployment. This rigid architectural approach faces significant challenges: it is ill-equipped to manage diverse, task-specific failures and often drifts out of sync with memory banks that change in both scale and structure as they evolve.

To overcome these constraints, we introduce MemPro, a framework designed for system-level evolution. Unlike traditional methods that adjust only the memory storage or prompt text, MemPro conceptualizes the entire MCR pipeline as an adaptable program. The system preserves a version tree of executable memory-system implementations. Within this structure, an Evolving Agent continuously evaluates promising iterations, identifies patterns in recurring failures, and generates enhanced child versions. This improvement process is driven by failure-mode-guided edit-debug refinement.

Our evaluations across LongMemEval, LoCoMo, HotpotQA, and NarrativeQA demonstrate that MemPro consistently surpasses strong baselines that are either static or limited to prompt-level evolution, often achieving superior results within just a few iterations. The system shows sustained improvement as evolution progresses, delivering an optimal balance between performance and cost. The source code for MemPro is publicly accessible at https://github.com/wanghai673/MemPro.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC