MemPro: Agentic Memory Systems as Evolvable Programs

Researchers introduce MemPro, an evolution framework that treats autonomous agent memory systems as adaptable programs rather than static pipelines. By iteratively diagnosing failures and refining the entire memory-construction-retrieval pipeline, MemPro outperforms fixed baselines on multiple benchmarks while maintaining computational efficiency.

MemPro addresses a fundamental limitation in current autonomous agent design: the rigidity of memory systems after deployment. Traditional agentic memory follows a construction-retrieval pipeline, but most systems only adjust the memory bank itself while leaving the surrounding architecture frozen. This inflexibility creates problems when agents encounter diverse failure modes or when memory structures must scale and evolve, often resulting in misalignment between the system's design and its operational needs.

The framework's innovation lies in treating the entire memory pipeline as an evolvable program rather than a static component. An Evolving Agent maintains a version tree of different memory-system implementations, progressively selecting versions that perform well, analyzing recurring failures, and generating improved iterations through guided refinement. This approach mirrors software engineering practices where systems adapt based on real-world performance data.

Experiments across LongMemEval, LoCoMo, HotpotQA, and NarrativeQA demonstrate consistent improvements over both static baselines and prompt-level evolution methods. MemPro achieves these gains within a few iterations while maintaining favorable performance-to-cost ratios, suggesting practical applicability without prohibitive computational overhead.

For the broader AI landscape, MemPro signals growing recognition that agent architectures require more sophisticated adaptation mechanisms than prompt engineering alone provides. The framework's open-source availability enables community iteration and refinement. Future implications extend to long-horizon AI systems in autonomous robotics, scientific discovery, and complex reasoning tasks where memory management directly impacts reliability and performance sustainability over extended operational periods.

• →MemPro evolves entire memory pipelines rather than just memory banks, addressing deployment rigidity in autonomous agents.
• →The framework uses failure-mode diagnosis to guide iterative refinement of memory-system implementations across a version tree.
• →Experimental results across four benchmarks show MemPro outperforms static and prompt-level evolving baselines within few iterations.
• →The system maintains favorable performance-cost trade-offs, indicating practical viability for real-world autonomous agent applications.
• →Open-source release enables broader adoption and community-driven improvements to agentic memory architectures.