Knowledge Injection via Addressed Conditional Memory

Testing on Qwen3.8-Flash-Next-FP8 and DeepSeek-V4.1-Flash shows that row-plus-projection updates achieve 28.7% natural-question target-string hits for post-hoc knowledge injection. While anchoring reduces model drift, it does not facilitate true entity-specific retrieval.