This report presents EMBER (Evolutionary MariaDB-driven Backend-integrated EPYC-accelerated Reflective-Agent), a production-ready, fully autonomous cognitive architecture in continuous operation since October 2025. The architecture runs entirely on non-accelerated CPU hardware — 16 dedicated cores of an AMD EPYC 9645 "Turin" — and orchestrates a 26-billion-parameter Mixture-of-Experts (MoE) core model (Gemma 4 26B) within an asynchronous, dual-context PHP backend paired with a persistent RAID-10 NVMe MariaDB as episodic long-term memory (LTM). We present three technical contributions: (1) a RAG-Lite system built on native FULLTEXT indices with Boolean-mode hardening for low-latency lore retrieval; (2) a probabilistically gated, context-isolated self-reflection pipeline with hard volatility hygiene; and (3) a vision-language closed-loop browsing system that streams base64-encoded screenshot frames with a coordinate-calibrated ghost cursor over a Server-Sent Events channel. We empirically characterize the hardware's inference profile and show that on CPU-only hardware the dominant latency driver is prompt mass, not model size — a finding that motivates the lore-gating and prompt-trimming mechanisms central to the architecture. We also report the Thinking-Bleed phenomenon: enforcing JSON grammar on a thinking-enabled model causes output to migrate to the internal thinking channel, a failure mode resolved by targeted per-call disabling of the thinking path.