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 retrieval. ↳ Instruct extracts structured facts and updates them across sessions. Together, they cover different failure modes. Redis says the combination matched or beat either approach alone across all six question types in their run. Cost matters too. Redis modeled the combined approach at about $0.07 per session, including answer-LLM costs. They also found that published benchmark results don't always reproduce under the same setup. A useful benchmark to check out if you're building agents with long-term memory. → fandf.co/4hdlNkg — Thanks to @Redisinc for sponsoring this post.

This is my website. Guess the stack.