MemTrace: Probing What Final Accuracy Misses in Long-Term Memory

In three linesMemTrace is a benchmark evaluating long-term memory in LLM agents across three dimensions: memory age, question type (current state, earlier state, trajectory), and evidence conditions. Testing 13 configurations, the study finds that evidence use is the primary bottleneck (10× more often retrievable than missing), not retrieval itself.

Summary generated by Claude — human-verified