Exabase M-1: Cheaper AI Memory Tops Benchmark

Exabase's memory engine, M-1, has achieved the highest reported score on the LongMemEval benchmark for conversational AI memory. It scored 96.4% accuracy at top-50 retrieval. What's interesting is that M-1 accomplished this using Gemini 3 Flash. This model is four to six times cheaper and faster than Gemini 3 Pro, which other leading systems on the benchmark used. Jonathan Bree, Exabase founder, highlights that M-1 was designed for production. Its memory architecture allows for better results with a cheaper, faster model. M-1 surpassed other systems like Mem0, Honcho, HydraDB, and Supermemory. The LongMemEval benchmark tests six capabilities, including recalling user facts and multi-session reasoning, across 500 questions. M-1's retrieval architecture draws on principles from episodic memory theory. The bottom line is that this development could mean more efficient and cost-effective AI memory systems for various applications.