@Letta_AI@OpenAI@AnthropicAI@claudeai "Recall benchmarks are saturated by agentic search" is a sharp call. What I'm more curious about is how you score "learn from experience." Is it later tasks going better, or whether what the agent writes to memory is actually right?