Reproduce “TrajectoryDB: A New Database for Agent Trajectories” 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
Yunjia Zheng, Juncheng Yang
TrajectoryDB frames agent execution histories as a distinct database workload rather than ordinary logs or documents. Its analysis of a month-long production trace shows why trajectory-native storage and querying are needed: sessions can resume after roughly 85 hours, 80% mutate their context, and branching sessions account for more than 80% of observed instances; the work proposes requirements and design directions, but does not present an implemented system or end-to-end benchmark.
Key insights
- Agent histories are better modeled as trees or DAG-like execution structures than as flat event streams: among 5,951 sampled sessions, forked sessions generated over 80% of all instances.
- Trajectory data has unusually irregular lifetimes and mutable context. The observed request inter-arrival time reached about 85.3 hours at the 99th percentile, while 80% of sessions changed context during execution.
- Semantic analysis should be treated as a database operator with measurable cost, selectivity, provenance, and cache dependencies—not as an external script appended after conventional query processing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.