Exabase Achieves State of the Art on BEAM Memory Benchmark at Every Scale, Using a Smaller, Cheaper Model

M-1 sets new records at 100K, 1M, and 10M tokens, outperforming every system on the leaderboard despite using a model 6x cheaper.

LONDON, July 28, 2026 /PRNewswire/ -- Exabase, a data infrastructure platform for AI agents, today announced that its memory engine M-1 has achieved the highest reported score on the BEAM benchmark at every evaluated scale: 76.9% at 100K tokens, 75.0% at 1M tokens, and 68.0% at 10M tokens, surpassing all previously reported systems.

As with Exabase's recent LongMemEval result (96.4%), M-1 achieved this using Gemini 3 Flash, a model up to six times cheaper, and faster, than the model (Gemini 3 Pro) used by every other leading system. M-1 also consumed approximately 20% fewer tokens per query than the next best system.

"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," said Jonathan Bree, founder of Exabase. "M-1 leads by the widest margin at the largest scale, which is exactly where recall architecture matters the most."

BEAM ("Beyond a Million Tokens"), introduced at ICLR 2026 by Tavakoli et al., is the most demanding public benchmark for conversational AI memory. It tests ten memory abilities across corpora ranging from 100K to 10M tokens. At 10M tokens, roughly equivalent to a year of daily agent conversations, fitting into any context window becomes physically impossible.

M-1's advantage widens at scale: the gap with previous leaders, Hindsight (73.4%, 73.9%, 64.1%), grows from 3.5 points at 100K to 3.9 at 10M, while the gap with Honcho (63.0%, 63.1%, 40.6%) expands from 13.9 to 27.4 points.

M-1 is now the only memory system to hold state-of-the-art results across both major conversational memory benchmarks at every evaluated scale, from 115K tokens on LongMemEval to 10M tokens on BEAM.

M-1's retrieval architecture was developed in collaboration with Hyperplane Labs (https://hyperplanelabs.ai), a European applied research laboratory focused on cognitive AI architectures.

M-1 is deployed in production, powering memory and search in products like Fabric (https://fabric.so), an AI workspace and personal data platform with over 300,000 users. The memory API is available to developers through the Exabase platform (https://exabase.io).

The full research paper, comparative results, and downloadable data are available at: https://exabase.io/research/exabase-achieves-state-of-the-art-on-beam-benchmark

About Exabase
Exabase is a context infrastructure platform for AI agents, providing memory, autonomous update, storage, search, and extraction capabilities. Exabase enables developers to build agents that persist, learn, and operate across sessions. Learn more at https://exabase.io.

About Hyperplane Labs
Hyperplane Labs is a European applied research laboratory focused on cognitive AI architectures. Learn more at https://hyperplanelabs.ai.

About Fabric
Fabric is a personal AI workspace and cloud computer where everything you save, write, capture, or record lives in one intelligent space. Learn more at https://fabric.so.

Contact:
Jonathan Bree
Founder, Exabase
[email protected]
+1 302-628-6654

SOURCE Exabase