Hayula Research Paper

We present Siyaq (Arabic: سياق — "context"), an open-source context engineering framework for AI agents that implements a six-stage pipeline — Perceive, Compress, Remember, Route, Govern, Reflect, Curate — to manage the complete lifecycle of agent context. Unlike existing approaches that treat context as passive input, Siyaq actively constructs, compresses, routes, and governs context to maximize utility while minimizing token consumption. Deployed on an Apple M2 Ultra (192GB) with 22+ local models and a 3-layer fallback chain (local → Qwen2.5:72b → OpenRouter), Siyaq achieves 60-90% token reduction through 95+ compression patterns, sub-50ms routing decisions, and zero cloud dependency for core operations. The framework powers 7 specialized routes across 4 Hayula ecosystem agents, serving 199K+ API requests monthly.