Research
RAG Serves the Superseded Value 36-38% of the Time on Real GitHub Fixes; a Supersession Memory Drives It to Zero
Temporal Validity on Real Software Histories (arXiv 2608.20685, Aug 21) extracts 130 clean atomic state transitions from 707 real GitHub issues in SWE-bench Lite and Verified, where a fix changes one identifiable value and the stale and current statements differ only in that value. Plain RAG scores 0.57-0.59 answer accuracy and, when forced to answer, returns the superseded value 36-38% of the time, with an LLM reranker failing to help while adding latency (~18s versus ~2.1s). A deterministic (subject, relation, object) supersession memory called MemStrata reaches 0.91 accuracy at RAG latency. The authors are explicit that only ~18% of real fixes are clean atomic transitions.
↳ Follow the thread