Researchers introduce Eywa, a memory system for AI agents that stores evidence separately from conclusions, improving reliability and auditability.

What happened: Eywa is a new memory architecture designed for AI agents that need to remember information across multiple sessions. Unlike existing systems that mix source evidence, extracted facts, and answers into one opaque path, Eywa stores immutable source evidence first, then derives facts from it, and retrieves context through a deterministic process without calling an AI model during retrieval itself.

Why it matters: AI agents that fail often do so for hard-to-diagnose reasons—missing evidence, unsupported extraction, stale information, or retrieval loss. By separating evidence from belief and recording every step, Eywa allows teams to audit where a wrong answer came from. The same memory substrate can also be evaluated across different answer models, from frontier to budget-constrained ones.

What to watch: On standard benchmarks, Eywa reached 90.19% judge accuracy on the LoCoMo C1-C4 split, 88.2% retrieval-sufficiency accuracy on LongMemEval-S, and 81.45% mean nugget score on BEAM. Full per-question artifacts—including questions, gold answers, model answers, retrieved context, and labels—are published publicly.