State-of-the-Art AI Memory Across All Scales

Hyperplane Labs · July 2026

In collaboration with Exabase, our research in cognitive retrieval architectures powers M-1, which has achieved the highest reported scores on both major AI memory benchmarks at every evaluated scale. The architecture draws on episodic memory theory (Tulving, 1972), reconstructive recall (Bartlett, 1932), and temporal context models (Howard & Kahana, 2002), treating memory as a multi-stage cognitive process rather than a search problem.

M-1 uses Gemini 3 Flash, a model 4-6x cheaper and faster than the Gemini 3 Pro used by every competing system, while consuming approximately 20% fewer tokens per query.

M-1 is the only memory system to hold state-of-the-art results on both LongMemEval and BEAM at every evaluated scale. Full results and methodology are published by Exabase: LongMemEval | BEAM.

The gap widens at scale

The most revealing pattern in these results is how the competitive gap changes with corpus size.

At BEAM-100K, M-1 leads the next system by 3.5 points. At 1M, the gap narrows to 1.1 points. At 10M, it widens again to 3.9 points. A third system on the leaderboard showed roughly flat performance from 100K to 1M, then collapsed to 40.6% at 10M, opening a 27.4 point gap.

The widening at the largest scale is the key finding. At 10M tokens, there is no context window large enough to compensate for weak retrieval. The corpus far exceeds any model's effective reasoning capacity. The only thing that determines the score is whether the retrieval architecture can surface the right information from a massive, noisy corpus.

If model scaling were converging on a solution to AI memory, you would expect the gap to narrow at larger scales as bigger models flex their capacity. The opposite happens. Our cognitive retrieval approach shows its greatest advantage at the hardest scale.

Four tiers of memory capability

BEAM tests ten distinct memory capabilities. Category-level analysis reveals four tiers of scale sensitivity, each telling us something about what current architectures handle well and where fundamental new approaches are needed.

The first tier is scale-invariant. Preference following, instruction following, summarisation, and abstention remain above 89% at all scales including 10M. These involve explicit, salient signals in the corpus, distinctive vocabulary and clear instructions that retrieval handles well regardless of corpus size. This tier is largely solved by current approaches.

The second tier shows graceful degradation. Event ordering and information extraction decline steadily. Event ordering drops from 83.6% at 100K to 67.5% at 10M. These require locating specific facts and reasoning about their relationships. The retrieval problem gets harder with scale but the challenge is quantitative, not qualitative. Incremental architectural improvements can continue to push these numbers.

The third tier shows moderate degradation. Contradiction resolution and temporal reasoning sit in a middle band, with temporal reasoning dropping from 66.3% to 58.8% across scales. These require understanding not just what was said but when, and whether it has been superseded. Temporal indexing and knowledge state tracking are the key capabilities here.

The fourth tier is catastrophic failure. Multi-session reasoning collapses from 44.7% at 100K to 9.6% at 10M. This category requires assembling fragments from multiple separate conversations to construct an answer. At 10M scale, the relevant fragments are scattered across a massive corpus with no shared surface features linking them. No existing system handles this well. It is an open research problem for the field.

Why multi-session reasoning breaks

The collapse at the fourth tier deserves particular attention, because it points to a fundamental limitation of retrieval-based architectures, including the one behind these results.

The upper tiers work because the retrieval signal is strong. Explicit preferences and instructions contain distinctive vocabulary that is easy to match against a query. Even specific facts, while harder to find in a large corpus, are stated plainly and can be located with good retrieval.

Multi-session reasoning breaks because there is no single retrieval signal. The answer exists only as an inference across multiple fragments that may share no surface similarity. A user mentioned a budget constraint in one conversation, discussed a vendor in another, and expressed a timeline preference in a third. No retrieval query, however well-decomposed, will reliably surface all three when they share no vocabulary, no temporal proximity, and no explicit connection.

This is not a problem that better retrieval can solve on its own. It requires a process that pre-links related memories, identifies latent connections, and creates composite representations that can be retrieved as units rather than assembled on the fly.

From retrieval to cognition

In cognitive science, the process that addresses this is consolidation. During sleep, the brain replays, reorganises, and integrates fragmented experiences into coherent knowledge structures (Walker & Stickgold, 2004; Stickgold & Walker, 2013). Consolidation does not wait for a query. It proactively restructures memory so that future retrieval can succeed in cases where it would otherwise fail.

The four-tier framework points to a clear research direction. Tiers 1 and 2 are well served by retrieval architecture, the work that contributed to M-1's results. Tier 3 requires better temporal reasoning and knowledge state tracking. Tier 4 requires active memory maintenance: consolidation, abstraction, and reorganisation processes that operate on the memory store itself, before any query arrives.

This is the focus of our next phase of research at Hyperplane Labs. The results presented here establish the baseline, demonstrating that cognitive retrieval architecture outperforms brute-force context at every scale. The open question is whether cognitive maintenance processes can address the failure modes that retrieval alone cannot.

The evidence from cognitive science is strong, the failure modes are well-characterised, and the evaluation frameworks exist to measure progress. The hardest problems in AI memory are now clearly defined. What remains is the research.

References

Bartlett, F.C. (1932). Remembering: A Study in Experimental and Social Psychology. Cambridge University Press.

Hong, K., Troynikov, A., and Huber, J. (2025). "Context Rot: How Increasing Input Tokens Impacts LLM Performance." Technical Report, Chroma.

Howard, M.W. & Kahana, M.J. (2002). "A distributed representation of temporal context." Journal of Mathematical Psychology, 46(3), 269-299.

Hsieh, C-P. et al. (2024). "RULER: What's the Real Context Size of Your Long-Context Language Models?" COLM.

Liu, N.F. et al. (2024). "Lost in the Middle: How Language Models Use Long Contexts." TACL.

Modarressi, A. et al. (2025). "NoLiMa: Long-Context Evaluation Beyond Literal Matching." ICML.

Stickgold, R. and Walker, M.P. (2013). "Sleep-dependent memory triage." Nature Neuroscience, 16(2), 139-145.

Tavakoli et al. (2025). "Beyond a Million Tokens." ICLR 2026.

Tulving, E. (1972). "Episodic and semantic memory." In Organisation of Memory, Academic Press.

Walker, M.P. and Stickgold, R. (2004). "Sleep-dependent learning and memory consolidation." Neuron, 44(1), 121-133.

Wang, D. et al. (2024). "LongMemEval: Benchmarking Long-Term Memory in AI Assistants."

arXiv:2601.15300 (2026). "Intelligence Degradation in Long-Context LLMs."