MemLens (64K) - Multi-Session Reasoning: leaderboard

Accuracy in percent on the MemLens 143 questions that combine evidence from several sessions, at a 64K-token interleaved image and text conversation history, with the evidence sessions hidden among distractor sessions (MemLens, 2026), judged by Qwen3-VL-235B. Higher is better.

Metric: Accuracy (%) on the multi-session reasoning (MSR, 143 questions) at a 64K-token history, interleaved multimodal multi-session conversation history with the evidence sessions hidden among haystack sessions, judged correct or incorrect by a Qwen3-VL-235B-A22B-Instruct judge; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 27 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Kimi K2.535.66#266
2Gemini 3.1 Pro (Preview)31.47#99
3GPT-5.427.27#135
4Qwen 3.5 122B A10B27.27#346
5GLM-4.6V23.78#710
6Gemma 3 27B (IT)22.38#1324
7Gemma 3 12B (IT)22.38#1887
8Qwen 3 VL 2B Instruct21.68#2367
9Qwen 3.5 27B20.98#375
10Qwen 3 VL 235B A22B Instruct20.98#532
11Phi-4 Multimodal Instruct20.98#4154
12Nemotron Nano 12B v2 VL20.98#2368
13Qwen 3 VL 30B A3B Instruct20.28#895
14Cosmos-Reason2-8B20.28#755
15Qwen 3 VL 8B Instruct19.58#1029

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=memlens-64k-multi-session-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-21.