Yiran Suo co-authored ContextWeave, a benchmark of whether remembered experience helps agents continue real workflows. It extends memory evaluation beyond recalling a fact from a long prompt.[1]
Start with ContextWeave →Contributions
Evaluate memory through completed work
ContextWeave reconstructs privacy-preserved workflows from 14 participants into executable tasks. The benchmark measures the resulting workspace and whether an agent follows the participant’s preferences. It also tests misleading recall, since remembered information can harm performance when an agent applies it to the wrong situation. Suo is a named co-author of this collaborative benchmark.[1]
Connect environment construction with agent training
Suo also appears in the Nex-N1 report’s contributor list. The report describes an ecosystem for constructing executable environments and using them in agent training.[3]
Tools, writing & talks
Writing, research & talks
Start here
ContextWeavearxiv.orgA benchmark of agent memory measured through continued office workflows and preferences.
Start with a workflow example and evaluation rubric. Running the benchmark requires familiarity with containerized agent evaluations.
Co-author[1]- Nex-N1 technical reportarxiv.org
A collaborative report on executable environments for agent training.
Co-author[3]
Tools
- ContextWeave implementation
github.comThe benchmark repository, with instructions for its tasks and evaluation.
Research collaborator[2]