Hayula Research Paper

We present TRUSTMEM-Hayula, an integrated approach combining TRUSTMEM's trustworthy memory consolidation with EvoTest's evolutionary self-improvement, deployed within the Siyaq context engineering framework. TRUSTMEM (arXiv:2606.25161) introduces a Memory Transition Verifier and preference-guided reinforcement learning that achieves 79% reduction in memory corruption by verifying transitions before consolidation. EvoTest (arXiv:2510.13220, ICLR 2026) introduces an Evolver Agent that analyzes execution episodes and autonomously revises the Actor Agent's configuration for continuous performance improvement. By integrating both into Siyaq's Remember and Reflect/Curate stages, we create a self-improving agent system where memory is not only trustworthy but continuously refined through evolutionary analysis. Our production deployment on Apple M2 Ultra (192GB) with Siyaq v2.3.0 demonstrates 43+ consecutive clean production cycles across a 4-tier memory hierarchy (ephemeral, working, long-term, archival) backed by SQLite FTS5 full-text search and the Hayula Gateway unified API endpoint.