Priors adds FrontierMemory v1.1 agent memory benchmark

Static trivia recall was never real agent memory, and folding FrontierMemory into custom weighted indices makes continual-learning capability a first-class model selection criterion alongside price and latency.

• –Unlike needle-in-a-haystack tests, Letta's FrontierMemory v1.1 evaluates how well a model "dreams" over 750 past agent trajectories to write structured memory files that boost downstream coding and knowledge-work pass rates
• –Integrating the dataset into Priors lets developers blend long-term memory synthesis directly into custom model rankings and subscription-adjusted cost comparisons rather than treating state management as an isolated side metric
• –Cost-versus-accuracy tradeoffs stand out starkly across the 17 tracked models: newly evaluated Gemini 4 Carbon reaches 71.1% at roughly $94 per dream (about a third of Opus 5.5's dreaming cost), while DeepSeek V4.1 Flash matches 71.1% as an open-weights contender
• –Steep drops from generalists like GPT-6.1 Sol (51.1%), Grok 4.7 (37.8%), and GPT-6 Luna (28.9%) underscore that strong reasoning benchmarks do not automatically translate into writing useful persistent context for agents