Reproduce “Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents” 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

Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz

The system checks candidate memories against a live database before saving them, using a separate after-task curator with read-only environment access. In a 40-question SQLite schedule with schema drift after question 20, probing raised pass rate from 70% to 73% versus memory-only curation and reduced task-agent queries from 5.6 to 4.7 per question; probe access and curator instructions changed together.

Key insights

  • Persistent memory is also an evidence-validation problem: useful records must be checked for transferability, scope, preconditions, and current validity rather than merely summarized from past trajectories.
  • Read-only probing is particularly valuable around environmental drift, such as database schema migrations, but it also helps validate file locations, workbook mappings, tool conventions, and reusable procedures in stable environments.
  • The best memories are actionable and scoped procedures, not isolated corrections or instance-specific answers; probing allows the curator to turn observations into executable guidance for later agents.

Summary by alphaXiv.

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