#London #England #AI memory #Exabase #M-1

Exabase's M-1 Engine Surpasses All Competitors in BEAM Benchmark Tests

In a groundbreaking development for artificial intelligence (AI), Exabase, a prominent data infrastructure platform, has unveiled that its memory engine, M-1, has achieved record-high scores on the BEAM benchmark at various scales. Notably, M-1 reached impressive scores of 76.9% at 100K tokens, 75.0% at 1M tokens, and 68.0% at 10M tokens, outclassing all other systems currently noted in the leaderboard.

Innovative Advancements in AI Memory

Exabase’s success with the M-1 memory engine is particularly remarkable given that it employs a model, namely Gemini 3 Flash, which is up to six times cheaper and faster than its competitors' models. This efficiency enables M-1 to operate with about 20% fewer tokens per query, a significant advantage in memory retrieval and processing.

Jonathan Bree, the founder of Exabase, emphasized the challenges at larger scales, stating, “At 10 million tokens, there’s nowhere to hide. The corpus doesn’t fit in any context window, and even if it could, context rot would degrade the results. The only path is real memory.” The M-1 system's architecture includes advanced retrieval mechanisms crafted in collaboration with Hyperplane Labs, a research facility specializing in cognitive AI architectures.

BEAM Benchmark: The New Gold Standard

Established at the ICLR 2026 conference by researchers Tavakoli et al., the BEAM benchmark is recognized as the most rigorous public evaluation for conversational AI memory. It effectively tests ten critical memory capabilities, spanning token counts from 100K to 10M. As the demand for more sophisticated memory systems grows, M-1 leads the way, significantly widening the gap against previous frontrunners such as Hindsight and Honcho.

For instance, M-1's lead over Hindsight has expanded from 3.5 points at 100K tokens to 3.9 points at 10M, while the divergence from Honcho increased from 13.9 to 27.4 points in the same comparative scales.

Real-World Applications and Availability

M-1 is not just a theoretical construct; it has proven its capabilities in production environments, powering key memory and search functions in various products, including Fabric, an AI workspace that has seen over 300,000 users. Developers are also granted access to the memory API through the Exabase platform, fostering innovation and application in diverse fields.

To explore the full range of capabilities and comparative results regarding the M-1 engine, interested parties can access the comprehensive research paper and downloadable data available through Exabase's official channels.

The Future of AI Memory Systems

The implications of Exabase’s advancements extend beyond mere performance metrics; they signify a transformative shift in how AI agents handle memory and context-awareness across sessions. The ability of M-1 to outperform competitors while utilizing a more cost-effective model could greatly impact the AI landscape, making sophisticated memory systems more accessible to developers and organizations alike.

With continued investment in enhancing memory infrastructure, Exabase is positioning itself at the forefront of AI technology. As AI systems evolve from simple applications to complex conversational agents, the need for robust and efficient memory solutions like M-1 will become paramount. This achievement not only sets a new benchmark for AI memory systems but also redefines what is possible in the realm of artificial intelligence.