Hayula Core: A Self-Improving Agent Platform with Skill-Based Architecture and Multi-Layer Memory

We present Hayula Core, a self-improving AI agent runtime built from 100% original Python code that replaces Hermes/OpenClaw as the central execution engine of the Hayula Agent OS ecosystem. Unlike existing agent frameworks that wrap monolithic LLM calls in fixed pipelines, Hayula Core introduces a skill-based architecture with 48 registered skills organized across 9 categories, a self-healing cron scheduler with production-proven automatic failure recovery, a four-layer memory system (ephemeral, working, long-term, archival), and a native Telegram bot integration for real-world deployment. The platform operates a self-improvement learning loop — agents evaluate, modify, and regenerate their own skill code based on execution outcomes, enabling continuous autonomous evolution. Deployed on an Apple M2 Ultra (192GB) and Linux infrastructure, Hayula Core drives 91 agents across 4 devices and 8 integrated frameworks with zero cloud dependency. We present the full architecture, analyze the 4-layer memory hierarchy with SQLite FTS5 and vector embedding support, document the self-healing cron scheduler's 43+ consecutive clean cycles, and compare the platform against LangChain, AutoGPT, and CrewAI across dimensions including architectural clarity, memory capabilities, skill extensibility, deployment footprint, and autonomous self-improvement. All code is released under the Hayula ecosystem at https://github.com/BinSaqban/hayula-core.