Eleven of twelve memory-system pairings failed to beat memory-off.

VibeMemBench separates two claims every memory pitch conflates. On 111 coding targets from 90 SWE-rebench V2 repositories with 3,634 history trajectories, directly injecting verified experience raised resolution on four of five held-out solvers by 1.1 to 4.5 points and cut agent steps on all five. But when four existing memory systems had to construct and retrieve that experience themselves, eleven of twelve solver/system pairings lost to the matched memory-off baseline. The experience is in the history. The retrieval layer is what's missing. (arXiv 2609.23570)