Letta Memory Benchmark - Core Memory Write (Hard): leaderboard

The Letta agent must write information into its core memory block through the memory tool, on the task's hard variant, over 100 items repeated 3 times; the score is the share done correctly.

Metric: Score (%) on 100 hard core-memory write 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 Sonnet 4 (20250514)100#388
2Claude Sonnet 4Thinking100#406
3GPT-4o Mini (2024-07-18)98.33#1511
4Gemini 2.5 Flash (Preview 05-20)97#423
5GPT-4o (2024-08-06)96.67#736
6Gemini 2.5 Flash (Preview 04-17)95.67#458
7Gemini 2.5 Pro (Preview 05-06)94.33#279
8O4 Mini93.33#356
9Claude 3.7 Sonnet (20250219)92#414
10Claude 3.5 Haiku (20241022)91.67#1601
11GPT-4.1 Nano78.67#2035
12Llama 4 Scout Instruct78.67#1579
13Llama 3.3 70B Instruct73#1379
14GPT-4.1 Mini70#781
15Llama 4 Maverick Instruct FP856.67#1010

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-write-hard · How It Works · Data refreshed daily, snapshot 2026-11-16.