MemShift separates noticing a changed fact from repairing memory

Five kinds of change

The benchmark changes a source record, deletes information, changes a personal definition, switches the model writing memories or exports memory between implementations. Synthetic worlds provide known answers before any narrative is generated, so scoring does not depend on another language model.

MemShift separates noticing a changed fact from repairing memory ↗

Detection alone leaves a gap

The authors evaluate four open-weight models and find that noticing a change helps only when the revised value is correct. Programs operating on structured records close much of the gap in their tests. This suggests checking invalidation and recomputation separately when evaluating memory.

MemShift separates noticing a changed fact from repairing memory ↗MemShift benchmark and reference memory ↗

What you can do

1. Compare the five controlled changes with failures your memory system needs to detect. Open reference ↗

Where the evidence stops

Abstract and repository read November 5. Synthetic records do not establish production privacy or reliability. No benchmark or model inference executed.

Sources read. Sources checked November 5, 2026 UTC. No third-party software, model inference or full workflow executed.

Sources & further reading