Maximem Synap is now SOTA on LongMemEval Benchmark

An update since we launched here last month.
Maximem Synap, our memory layer for AI agents, now scores 95.6% on LongMemEval across all 500 questions. That is SOTA for agent memory, ahead of every published score from Mastra, Hindsight, Mem0 and Zep.

Two things matter to us more than the rank.

• Synap gets there with Gemini 3.8 Flash, a fast and inexpensive answer model, so the memory is doing the work rather than a frontier model. And it hands that model about 7,400 tokens per question out of a roughly 115,000-token history, 4x to 6x less than the next two systems, which is what keeps cost and latency low in production.

• Every answer and judge verdict from the run is public, and the eval harness is open source if you want to run the same test on your own memory system.

Results and the full scorecard: https://www.maximem.ai/blog/synap-number-one-longmemeval

If you are running Maximem Synap today, I would like to hear which conversations your agent still gets wrong. That is where we are spending the next month.