Eywa: Local-first memory for AI agents, with a receipt for every fact (arxiv.org)

Researchers have introduced Eywa, a revolutionary memory architecture designed for AI agents that require persistent, retrievable, and auditable memory capabilities. Unlike conventional memory systems that intertwine various layers of evidence, facts, and retrieval processes, Eywa emphasizes a "provenance-grounded" approach where source evidence is stored first, ensuring that each fact derived can be validated against its original context. This method allows AI agents to retrieve memory more accurately and transparently, enhancing their ability to provide reliable answers and reducing the complexity involved in diagnosing errors. The significance of Eywa lies in its ability to maintain high accuracy under controlled retrieval conditions. For instance, it achieved a remarkable 90.19% judge accuracy on the LoCoMo C1-C4 benchmark and an 88.2% retrieval-sufficiency accuracy on LongMemEval-S. By employing deterministic multi-route read paths that minimize calls to large language models during retrieval, Eywa positions itself as a robust solution for long-term memory management in AI systems. The release of detailed data artifacts related to this research further contributes to the community's understanding and potential applications of effective memory structures in AI and ML.