[arXiv]score: 0.20
Eywa Agent Memory Architecture Separates Evidence from Belief with Zero-LLM Retrieval
June 1, 2026
Eywa stores immutable source evidence before deriving facts, validates extractions against typed signals, and retrieves memory context via a deterministic multi-route path with zero LLM calls inside retrieval. The architecture separates retrieved context from answer instructions, enabling the same memory substrate to be evaluated across different answer models.
cs.CL
HOW THIS AFFECTS YOU
●
builderThe zero-LLM retrieval path and provenance separation are directly applicable to production agent systems where auditability, stale-state bugs, and GDPR-style erasure requirements are real constraints.
●
researcherThe evidence-before-belief design and frozen artifact evaluation setup offer a rigorous framework for isolating memory failure modes from answer model behavior.