LongMINT: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems

Researchers introduced LongMINT, a benchmark for testing memory-augmented AI agents in long-horizon, interference-heavy settings across multiple domains.

LongMINT includes 15.6K QA pairs built from very long, frequently updated contexts and evaluates both single-target recall and multi-target aggregation.

Across seven tested systems, performance was generally poor, with retrieval and memory construction emerging as the main bottlenecks.

The benchmark exposes major weaknesses in current agent memory systems and offers a more realistic way to measure recall and reasoning over evolving information.