New benchmark VetoBench tests if AI agents remember rejected tech decisions

A developer has introduced VetoBench, a small reproducible benchmark designed to test whether AI agents avoid re-proposing previously rejected technical approaches, not just whether they recall accepted ones. The benchmark uses 24 synthetic engineering decisions, 10 of which include structured rejection records with documented reasons such as failed spikes or rolled-back migrations. Without any memory, the tested agent (Claude Haiku 4.5) re-proposed rejected solutions in 80–90% of scenarios; providing vetoes directly in context dropped that rate to 0%. Memory platform Mem0 performed reasonably well with a 0–20% violation rate, but an audit of retrieved contexts revealed that in 38% of test cases the rejection record was simply absent from what the system retrieved. The findings suggest that storing and reliably surfacing veto decisions — not just accepted choices — is a critical and largely overlooked gap in current AI agent memory systems.