Passport Control
An AI agent memory-writing benchmark published on Kaggle. It comprises 661 checks designed to evaluate large language models' capabilities in the memory-writing phase, covering four categories: retention, rejection, temporal annotation, and modification.
Timeline (last 90 days)
Passport Control is a benchmark published on Kaggle that specifically tests the write phase of AI agent memory, comprising 661 checks.
Passport Control groups its 661 checks into four families: Keep, Refuse, Date, and Change.
All test dates used in Passport Control extend from late 2026 into 2027, beyond the training cutoff of all models tested, in order to prevent memory cheating.
The benchmark tested 18 models from six labs in a single run, covering Anthropic, OpenAI, Google, DeepSeek, Qwen, and GLM.
The total cost of a full write run across all 18 models was only $15.11, with scoring determined entirely by code.
11 of the 18 models correctly answered at least 96% of memory read questions, but the gap widened significantly on the memory write task (maintaining the passport).
The bottleneck in agent memory lies in the write path, yet most memory benchmarks (such as LongBench and Needle-in-a-Haystack) only test long-context retrieval and reasoning, implicitly assuming that memory has already been written correctly.
Setting a default expiration period for each category of facts is a low-cost engineering patch that can fix the memory non-expiry problem without requiring stronger reasoning capability from the model.
All Facts (8)
Passport Control is a benchmark published on Kaggle that specifically tests the write phase of AI agent memory, comprising 661 checks.
50%UnverifiedPassport Control groups its 661 checks into four families: Keep, Refuse, Date, and Change.
50%UnverifiedAll test dates used in Passport Control extend from late 2026 into 2027, beyond the training cutoff of all models tested, in order to prevent memory cheating.
50%UnverifiedThe benchmark tested 18 models from six labs in a single run, covering Anthropic, OpenAI, Google, DeepSeek, Qwen, and GLM.
50%UnverifiedThe total cost of a full write run across all 18 models was only $15.11, with scoring determined entirely by code.
50%Unverified11 of the 18 models correctly answered at least 96% of memory read questions, but the gap widened significantly on the memory write task (maintaining the passport).
50%UnverifiedThe bottleneck in agent memory lies in the write path, yet most memory benchmarks (such as LongBench and Needle-in-a-Haystack) only test long-context retrieval and reasoning, implicitly assuming that memory has already been written correctly.
50%UnverifiedSetting a default expiration period for each category of facts is a low-cost engineering patch that can fix the memory non-expiry problem without requiring stronger reasoning capability from the model.
50%