Memora benchmark shows AI agent memory goes stale

Researchers from Arizona State and the University of Arizona released Memora (paper "From Recall to Forgetting", Findings of ACL 2026), a benchmark that scores not just recall but the penalty for using outdated facts. Across simulations of up to 1,991 sessions, memory agents nearly double LLMs on recall (119.5 vs 65.7 FAMA) but underperform on recommendations and accumulate more stale-memory errors as history grows.

• Memora simulates 10 personas across 155, 615 and 1,991 sessions of ~16 turns
• Quarterly setting averages ~1,171 memory operations: 63% additions, 18% updates, 19% deletions
• Memory agents score 119.5 vs 65.7 for LLMs on recall under FAMA
• Stale-memory penalty for agents grows from 18.2 to 29.5 points