Letta Memory Benchmark - Archival Memory Read: leaderboard

The Letta agent must search its archival memory to answer a question, over 100 items repeated 3 times; the score is the share answered correctly. It is the retrieval part of the Letta Memory Benchmark.

Metric: Score (%) on 100 archival-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
1Claude 3.5 Haiku (20241022)100#1601
2GPT-4.1 Mini99.67#781
3Gemini 2.5 Pro (Preview 05-06)98.67#279
4Claude Sonnet 4Thinking98.33#406
5GPT-4.1 Nano94.33#2035
6GPT-4o (2024-08-06)92.33#736
7Gemini 2.5 Flash (Preview 05-20)92.33#423
8Gemini 2.5 Flash (Preview 04-17)89#458
9GPT-4o Mini (2024-07-18)82.67#1511
10Claude Sonnet 4 (20250514)81.33#388
11Llama 4 Scout Instruct64.33#1579
12Llama 4 Maverick Instruct FP849.33#1010
13Claude 3.7 Sonnet (20250219)47#414
14Llama 3.2 Instruct Turbo (3B)12.67
15Llama 3.3 70B Instruct8.33#1379

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-archival-memory-read · How It Works · Data refreshed daily, snapshot 2026-11-16.