MemAlign: Building Better LLM Judges From Human Feedback With Scalable Memory

Companies increasingly rely on large language model judges to evaluate AI systems, but these judges often miss domain-specific quality standards that human experts prioritize. MemAlign solves this using a dual-memory system that learns from small amounts of natural language feedback rather than requiring hundreds of labeled examples. The system stores general principles in semantic memory and specific examples in episodic memory, enabling it to adapt in seconds while costing 10-100 times less than existing tools, achieving 90%+ accuracy on tested examples.