ExperienceIndex: Artifact-Grounded Memory
Abstract Overview
This paper introduces ExperienceIndex (ExpIdx), a memory layer for AI agents that stores and reuses artifact-grounded experience extracted from prior reasoning traces. Instead of focusing on user profiles or abstract reasoning strategies, ExpIdx treats artifacts as the anchor of memory and represents experience in two forms: single-artifact summaries of prior contributions and artifact-pair relationships that capture structural connections. The system is implemented as lightweight middleware that retrieves relevant experiences for a new task and injects them into both the model prompt and the search framework output. The authors evaluate ExpIdx across multiple domains, modalities, and agentic search frameworks, including files, documents, codebases, and tables.
Novelty
The distinctive contribution is an artifact-centered memory mechanism for agentic systems, designed to preserve and reuse knowledge about specific artifacts encountered in earlier tasks. Its combination of single-artifact and artifact-pair experiences, along with middleware-based retrieval and injection, differentiates it from prior memory approaches centered on user profiles, factual attributes, abstract reasoning patterns, or offline corpus enrichment.
Results
Across the reported evaluations, ExpIdx consistently improves answer quality while lowering online cost, with gains of up to 11.0 score points and up to 50.5% reduction in online dollar cost relative to the underlying agentic solutions. On average across evaluation models, it achieves the highest answer score and the lowest online cost among the compared methods, while also reducing indexing cost and storage footprint substantially relative to the offline enrichment baseline. The paper also reports cross-task transfer from Spider to MMQA and a teacher-student setting in which experiences from a stronger model help a weaker model reach comparable performance at lower cost.
Key Points
- ExpIdx stores reusable artifact-grounded experience from prior reasoning traces in two complementary forms: single-artifact contributions and artifact-pair relationships.
- The method operates as lightweight middleware that refines retrieved artifacts and augments the model context, making it compatible with multiple search frameworks and task settings.
- Empirical results indicate consistent quality and efficiency improvements, plus evidence of cross-task generalization and teacher-student transfer over shared artifact corpora.