Supermemory Claims 99% SOTA on Agent Memory Benchmark — Critics Question the Cost
A new agent memory system called Supermemory is claiming near-perfect scores on LongMemEval, but skeptics point out it achieves this by running 18 agents in parallel for a single memory recall, raising hard questions about benchmark gaming versus real-world viability.
Supermemory's announcement, shared by @DhravyaShah, generated significant buzz with its claim of ~99% accuracy on the LongMemEval benchmark — a standard test for how well AI systems can retrieve and utilize information from long conversation histories. "AI agents will now remember EVERYTHING," the post declared. The system uses what appears to be an adaptive semantic memory retrieval architecture that the team calls ASMR.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.