Letta introduces FrontierMemory to score how memory from past agent attempts helps solve later tasks

FrontierMemory asks agents to generate memory from earlier task transcripts, then measures whether it helps solve later tasks, Letta explains. Its published tables give Opus 5.5 and GLM-5.3 scores of 73.3% among models, while Letta Dream leads the separate harness table at 73.3%. Those are Letta’s results for this benchmark, not a general ranking of agent ability. Longer-horizon tasks remain a future plan.