Gaurav Dadhich

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/syna...

At Maximem, a frontier agent with multi-level judges builds a custom context architecture for each agent that uses Maximem. This ensures that you are not getting a generic memory system but one that is tailored for your agent's needs.