Letta Memory Benchmark: leaderboard

Each model runs as a Letta agent with memory tools on four tasks: archival memory read, core memory read, core memory update and a hard core memory write, each over 100 items repeated 3 times; the score is the published mean of the four task scores. Unlike Letta's Context-Bench, which grades file and skill use, it tests the agent's own memory blocks.

Metric: Average score (%) over four Letta memory tasks, 100 items each, 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 Sonnet 4Thinking92.08#406
2Claude Sonnet 4 (20250514)91.25#388
3GPT-4o (2024-08-06)90.5#736
4GPT-4o Mini (2024-07-18)87.83#1511
5Gemini 2.5 Flash (Preview 05-20)87.58#423
6Gemini 2.5 Pro (Preview 05-06)84.92#279
7GPT-4.1 Mini80.17#781
8Claude 3.7 Sonnet (20250219)78.58#414
9Claude 3.5 Haiku (20241022)70.5#1601
10Gemini 2.5 Flash (Preview 04-17)67.92#458
11O4 Mini66.5#356
12Llama 4 Maverick Instruct FP861.25#1010
13Llama 3.3 70B Instruct55.67#1379
14GPT-4.1 Nano47.75#2035
15Llama 4 Scout Instruct47#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 · How It Works · Data refreshed daily, snapshot 2026-11-16.