Exabase M-1 Tops LongMemEval Benchmark Using Smaller, Cheaper Model

Exabase that its memory engine M-1 achieved the highest reported score on the LongMemEval benchmark for conversational AI memory. The system scored 96.4 percent accuracy at top-50 retrieval, surpassing previous leaders such as Mem0, Honcho, HydraDB, and Supermemory.

M-1 reached this result using Gemini 3 Flash, a model four to six times cheaper and faster than Gemini 3 Pro, which powered competing systems. The benchmark evaluates conversational memory across 500 questions and over 115,000 tokens, testing recall, reasoning, and knowledge update capabilities.

The M-1 retrieval architecture was developed with Hyperplane Labs and is based on principles from episodic memory theory and reconstructive recall. The system is already in production, supporting memory and search in the Fabric AI workspace, which has more than 300,000 users. Developers can access the memory API through the Exabase platform.