WhenLoss: Diagnosing Write and Retrieval Bottlenecks in Long-Context Memory Systems

In three linesWhenLoss introduces a diagnostic protocol to identify bottlenecks in long-context memory systems. Expected Predictive Compression (EPC) uses an LLM to anticipate future questions and preserve minimal evidence at write time. On LongMemEval (500 questions), EPC achieves 0.49 CSM score vs 0.44 for strongest baseline, reducing write-side gap to 0.04.

Summary generated by Claude — human-verified