ContextBench V2: Realistic Memory Evaluation for Letta Agents

Task-type scores reveal distinct strengths in using and maintaining memory.

Each model's mean score plotted against what a run costs, with the Pareto frontier marking the best trade-offs available at each price.