MemTrace: New Benchmark for LLM Long-Term Memory Accuracy

What Changed

Introduces knowledge-point-based evaluation instead of question-based aggregation.

Why It Matters

This research shifts the focus of memory optimization from increasing storage capacity to improving reasoning over retrieved context. Developers should prioritize better evidence synthesis logic in their RAG pipelines.

What To Do Next

Audit your RAG pipeline to see if the model is correctly synthesizing retrieved evidence, rather than just focusing on improving retrieval recall.