Systems with large language models (LLMs) and persistent memory face a largely unaddressed problem: Knowledge Corruption — the gradual degradation of stored knowledge through epistemically invalid write operations. Unlike hallucination, which originates within the model, Knowledge Corruption occurs when correctly retrieved information is epistemically compromised at the system level. We present a taxonomy of five Knowledge Corruption patterns observed in a production LLM knowledge system and propose the Validation Gates Framework — a pipeline of 6+1 epistemic validation layers that operate at write time. We implement six gates in BrainDB, an SQLite-based knowledge system with approximately 5,000 entries, including a Cross-Model Validation prototype. Evaluation across 41 independent runs with 5 validator models from 4 providers (Anthropic, OpenAI, Google, and a local 3B-parameter model) shows an overall detection rate of 92% (σ = 0%) — 100% for Source Confusion, Confidence Decay, Stale Override, and Temporal Consistency — with 0% false positives across 8 clean control scenarios for 4 of 5 models.