Reproduce “STARK: Moving towards Personalized Agents That Truly Learn and Remember” 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

Aryan Jain, Abhishek Jha, Anish Nediyanchath, Promod Yenigalla

STARK separates interaction episodes from durable behavioral signals, such as preferences and constraints, links signals to supporting evidence, and infers candidates from users with shared patterns. On LongMemEval information extraction, it scores 4.38/5 versus MAGMA’s 2.90/5, using Claude Sonnet 4.6, a GPT-5.1 judge, and equal 4,000-token retrieval budgets. Transfer is limited to shared-domain users, and inferred signals may be wrong.

Summary by alphaXiv.

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