M-1 achieves state-of-the-art on BEAM at every scale, with a smaller model and fewer tokens

M-1 (Mneme-1), now holds the top score on both major conversational memory benchmarks, LongMemEval and BEAM, at every evaluated scale. No other memory system has done this.

BEAM tests what happens when memory corpora get genuinely large: 100K, 1M, and 10M tokens. At 10M tokens, context windows cannot hold the corpus. The only path to a good score is retrieval that actually works.

M-1 scored 76.9% on BEAM-100K, 75.0% on BEAM-1M, and 68.0% on BEAM-10M, leading every other reported system at all three scales. As with our LongMemEval result, we used Gemini 3 Flash while every other system used Gemini 3 Pro, which is four to six times more expensive. M-1 also consumed approximately 20% fewer total tokens per query than the next best system.

The competitive gap widens at the hardest scale: M-1 leads by 3.5 points at 100K and by 3.9 points at 10M. The gap getting larger when retrieval conditions are most challenging suggests that the architectural advantage becomes more pronounced, not less, as scale increases.

The consistency across different benchmarks and scales reinforces the finding from our earlier evaluation: retrieval architecture determines memory system quality more than model scale does. As context windows continue to grow, the case for precise retrieval only gets stronger, because a larger window gives you a larger container of text that the model cannot reliably reason through.

There is clearly more work to do, particularly on multi-session reasoning at extreme scale, which remains an open challenge for every memory system in the field. We will continue to share results as we improve.

Read the full research paper, methodology, and results here.