Reproduce “Governed Memory: A Production Architecture for Multi-Agent Workflows” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hamed Taheri

An architecture called "Governed Memory" provides a unified infrastructure for multi-agent workflows in enterprise AI, addressing challenges like fragmented knowledge and inconsistent policy application. This system achieved 99.6% overall fact recall, reduced token usage by 50.3% through progressive context delivery, and demonstrated 74.8% accuracy on the LoCoMo benchmark while ensuring robust governance and zero cross-entity data leakage.

Key insights

  • The dual memory model is complementary, effectively capturing distinct information types (e.g., long-tail insights in open-set memory and typed values in schema-enforced memory) that would otherwise be lost.
  • Progressive context delivery significantly optimizes token usage by injecting only new or relevant content into agent context windows, leading to substantial reductions in computational cost and improved model attention.
  • AI-assisted schema authoring, write-time quality gates, and automated schema refinement are critical for maintaining high memory extraction quality and ensuring the robust enforcement of organizational policies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.