Letta Memory Benchmark - Core Memory Read: leaderboard

The Letta agent answers from facts held in its in-context core memory block, over 100 items repeated 3 times; the score is the share answered correctly. Unlike the archival read task, the facts are already in context.

Metric: Score (%) on 100 core-memory read items, 3 repeats, Letta agent framework. Source: raw.githubusercontent.com. Saturation forecast: Estimated already saturated. 18 models tracked. Updated 16 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-4o (2024-08-06)100#736
2O4 Mini100#356
3Claude Sonnet 4 (20250514)100#388
4O3 Mini100#544
5Gemini 2.5 Pro (Preview 05-06)100#279
6GPT-4.1 Mini99.67#781
7GPT-4o Mini (2024-07-18)99.67#1511
8Gemini 2.5 Flash (Preview 05-20)99.67#423
9Claude 3.7 Sonnet (20250219)99#414
10Claude Sonnet 4Thinking97.67#406
11Llama 3.3 70B Instruct97#1379
12Llama 4 Maverick Instruct FP892.33#1010
13Claude 3.5 Haiku (20241022)62#1601
14Gemini 2.5 Flash (Preview 04-17)37.67#458
15Llama 4 Scout Instruct31.33#1579

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=letta-memory-benchmark-core-memory-read · How It Works · Data refreshed daily, snapshot 2026-11-16.