Evaluate agent memory on four cognitive competencies with MemoryAgentBench

arXiv·medium signal

MemoryAgentBench grounds memory evaluation in cognitive science and probes four distinct competencies — accurate retrieval, test-time learning, long-range understanding, and selective forgetting — by feeding inputs incrementally as time-ordered dialogue chunks across architectures (MIRIX, MemGPT, long-context, and RAG agents). The key insight for builders: 'remembers facts' is only one of four axes, and selective forgetting is usually the neglected one. Score your own agent's memory layer against all four so you catch failures like hoarding stale context or failing to update on new information, not just retrieval misses.