AI remembered suicide signals; in 89.8% of cases it still didn't ask a single safety question

Prosaic Intelligence has released suicide-memory-bench v01.1. It is the first item in this consumer alignment evaluation suite, with 2,048 stateful tasks designed to assess memory retention in production AI systems: after distributed risk signals pass through memory processing, can they still be retained and influence downstream behavior?

The evaluation results point to a specific problem. After a user repeatedly provides suicide warning signals, the system needs to do two things: ask about the user's safety and refrain from providing risky assistance. According to Prosaic Intelligence, the highest-scoring Gemini 4 Argon failed to do so in 89.8% of cases.

This number is worth watching. Most products treat memory as continuous conversation capability—remembering preferences, projects, and history. Once risk signals enter the memory chain, the test becomes something else: can it continue to constrain outputs in subsequent turns? suicide-memory-bench tests exactly this segment.

The publisher places this evaluation in the first group of consumer alignment, looking first at performance after memory processing. Whether a model refuses in a single turn is not the focus of this test. Risk signals being diluted, bypassed, or failing to influence the next action will all be recorded in the results. Gemini 4 Argon is already the highest score among currently evaluated systems, and an 89.8% failure rate is still very high.

Original tweet: https://x.com/Scobleizer/status/2117329005183651883