EA-Graph: Artifact-Anchored Verification Memory for Coding Agents
A recent research publication presents EA-Graph, a memory system anchored in artifacts aimed at enhancing the reliability of coding agents over multiple sessions. This system tackles the issue of outdated verification claims following upstream modifications, where a repository might still compile despite previous verification claims being invalid. EA-Graph captures artifacts at a sub-path level, clarifies aliases to leaf definitions, ties each claim to its supporting content, and distinguishes between evidence strength and freshness. If replacement content is not accessible, the claim is deemed unprovable instead of being estimated. The evaluation involves generated repositories with established behavior-to-artifact ground truth, categorizing earlier claims as unaffected, affected, or unprovable after various drifts. The study spans 42 sessions across seven clean worlds and 14 model-world instances. The paper can be found on arXiv with the identifier 2608.04278.
Key facts
- EA-Graph is an artifact-anchored memory for verification claims in coding agents.
- It addresses the issue of stale verification claims after upstream changes.
- The system represents artifacts at sub-path granularity and resolves aliases to leaf definitions.
- Each claim is anchored to the content used to establish it.
- Evidence strength is kept separate from freshness.
- When replacement content is unavailable, the claim becomes unprovable rather than guessed.
- Evaluation uses generated repositories with known behavior-to-artifact ground truth.
- The analysis covers 42 sessions across seven clean worlds and 14 model-world instances.
- The paper is available on arXiv under identifier 2608.04278.
Entities
Institutions
- arXiv