AI agents are only as useful as their memory. Redis just published a benchmark of agent memory systems using LongMemEval, a benchmark with 500 questions across multi-session conversations. Their Remis + Instruct approach reached 86.5% task-averaged accuracy, the highest result among the production systems they reproduced on a common GPT-4o backbone. The interesting part? It combines two different approaches: ↳ Remis keeps raw conversation evidence using hybrid semantic + lexical…