mnemos ships persistent memory, adds verifier

The strongest part here is the eval shape: same prompt, same model, memory toggled on and off, plus a separate capture test for whether agents actually retain corrections. That makes this feel more credible than a typical memory-layer launch post, even if the sample size is still small. The 5/5 vs 0/5 results on session-start and protocol-convention cases suggest mnemos mainly helps where the model lacks project-specific priors, while the 5/5 vs 5/5 results on common best-practice cases are a good sign it is filling gaps without disturbing what the model already knows. The jump from 7% to 53% on correction capture is the more important result, because write-side memory is usually where these systems fail. The UserPromptSubmit hook is the key architectural move: it turns correction detection into a non-skippable boundary event instead of a voluntary follow-through. The ceiling still matters: 53% means prompt framing alone is not enough, and the remaining misses look like task prompts that bury the correction too deeply. This is useful benchmark work, but n=5 per scenario is not enough to claim broad generalization yet.