Redis Agent Memory benchmark report

Key highlights:

• Remis + Instruct scored 86.5% task-averaged accuracy, the top result among production systems we reproduced on GPT-4o.

• It delivered accuracy effectively level with the best managed service we tested, which used a stronger answer model.

• At about $0.07 per session, Remis + Instruct cost less than one-third as much as other systems we reproduced above 80% accuracy.

• Published results differed from what we measured by as much as 28 percentage points, in both directions.

• The methodology is public, so you can reproduce the benchmark and test it against your own workload.