Exabase Achieves State-of-the-Art on BEAM Memory Benchmark with Smaller Model
Key Takeaways
- ▸Exabase's M-1 system achieved state-of-the-art on BEAM across all scales (100K–10M tokens), beating previous leaders by up to 27.4 points at 10M scale
- ▸Performance achieved using smaller Gemini 3 Flash model instead of larger Gemini 3 Pro, demonstrating efficiency gains from better architecture rather than model size
- ▸System shows strong performance on preference following, instruction following, and summarization (>90% at all scales) but struggles with multi-session reasoning at extreme scales (9.6% at 10M)
Summary
Exabase, a startup focused on AI memory systems, announced state-of-the-art results on BEAM, the hardest AI memory benchmark, across all scales up to 10 million tokens. The company's M-1 system scored 76.9% at 100K tokens, 75.0% at 1M tokens, and 68.0% at 10M tokens—outperforming previous leaders Hindsight and Honcho by 3.5 to 27.4 percentage points depending on scale. Critically, Exabase achieved these results using Google's smaller Gemini 3 Flash model, while competitors relied on the larger Gemini 3 Pro.
The BEAM benchmark tests ten memory abilities including contradiction resolution, event ordering, and instruction following, with the 10M token scale roughly equivalent to a year of daily chat interactions. At such extreme scales, traditional context-stuffing approaches fail, requiring fundamentally better recall mechanisms. Exabase's system also demonstrated 20% better token efficiency per query compared to the next-best system.
With these results combined with previous state-of-the-art on the LongMemEval benchmark, Exabase is now the only system holding top rankings across both major memory benchmarks at every scale from 115K to 10M tokens. The company published full methodology and results in a research paper, providing transparency into how M-1 achieves superior recall through architectural improvements rather than brute-force model scaling.
- M-1 consumes 20% fewer tokens per query than competing systems, indicating improved retrieval efficiency
- Exabase now holds SOTA on both BEAM and LongMemEval benchmarks, the two major AI memory evaluation frameworks
Editorial Opinion
This result is significant not because it merely tops a benchmark, but because Exabase achieves it with a smaller, more efficient model. In an era where scaling dominates AI progress, demonstrating superior recall through architecture is a meaningful contribution—and suggests the memory problem may be solvable without just throwing more compute at it. The 10M-token scale is notably grueling (past any model's context window), making this a genuine test of retrieval rather than context awareness. The weaknesses at multi-session reasoning reveal this isn't a solved problem, but Exabase's systematic approach to memory architecture deserves attention from anyone building persistent AI systems.
