模型矩阵
我们追踪的每个平台，汇总于一张信息密集的表格：个性化机制、各平台能力，以及便携档案的适配程度。点击列标题排序。功能数据均附有各条目的审核日期，无虚构评分。
| 平台 | Profile 匹配度 | 自定义指令 | 记忆 | Files | 网络搜索 | API | Free tier | 个性化说明 |
|---|---|---|---|---|---|---|---|---|
| ChatGPT OpenAI | excellent | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Yes, Custom Instructions (persistent), plus per-Project instructions |
| Claude Anthropic | excellent | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Yes. Project instructions, personal preferences, and custom Styles |
| OpenClaw Open source | excellent | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | Yes, per-agent workspace files (AGENTS.md, SOUL.md, USER.md) act as persistent… |
| Poe Quora | excellent | ✓ | ◐ | ✓ | ◐ | ◐ | ✓ | Yes. Every custom bot IS a persistent instruction set |
| Gemini Google | good | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Yes, Saved Info (persistent facts/preferences) and per-Gem instructions |
| Grok xAI | good | ✓ | ◐ | ✓ | ✓ | ✓ | ◐ | Yes, custom instructions in settings on current versions |
| Perplexity Perplexity AI | good | ✓ | ◐ | ✓ | ✓ | ✓ | ✓ | Yes, profile personalization prompt; per-Space instructions |
| GitHub Copilot GitHub (Microsoft) | good | ✓ | ◐ | ◐ | ◐ | ◐ | ◐ | Yes, repository instruction files (e.g.github/copilot-instructions.md) and personal… |
| Cursor Anysphere | good | ✓ | ◐ | ◐ | ✓ | ✕ | ◐ | Yes. Rules (global “Rules for AI” + per-project rules files) |
| Windsurf Windsurf | good | ✓ | ✓ | ◐ | ✓ | ✕ | ◐ | Yes, rules files + memories |
| Microsoft Copilot Microsoft | workable | ◐ | ◐ | ✓ | ✓ | ◐ | ✓ | Limited, persistent custom-instruction support varies; paste per session or use Copilot… |
| Le Chat Mistral AI | workable | ◐ | ◐ | ✓ | ✓ | ✓ | ✓ | Varies, agent-style customization via the developer platform; paste per chat in the… |
| Gemini Notebook (formerly NotebookLM) Google | workable | ✕ | ◐ | ✓ | ◐ | ✕ | ✓ | Notebook-level instructions on current versions; your profile can also live as a source |
| Meta AI Meta | basic | ✕ | ◐ | ◐ | ✓ | ◐ | ✓ | No dedicated persistent instruction box on most surfaces. Paste per conversation |
| DeepSeek DeepSeek | basic | ✕ | ✕ | ✓ | ✓ | ✓ | ✓ | No persistent instruction box. Paste per conversation |
| Midjourney Midjourney | different | ✕ | ◐ | ◐ | ✕ | ✕ | ◐ | Not text-profile based, visual personalization/style references instead |
| ElevenLabs ElevenLabs | different | ◐ | ◐ | ◐ | ✕ | ✓ | ✓ | For voice agents: system prompts supported, your profile applies there |
| Runway Runway | different | ✕ | ✕ | ◐ | ✕ | ✓ | ◐ | Not applicable |
✓ 支持 · ◐ 部分支持或因情况而异 · 不提供。本表为定性设计：相关情况变化频繁，每个平台页面均注明了其条目的最后审阅时间。
在找实时基准测试评分？
这是一项不同且难度更高的任务，LiveBench 在这方面做得很出色：提供抗污染的评测得分，涵盖推理、编程、数学、语言、数据分析与指令遵循（6 个类别共 18 项任务），持续更新。
livebench.ai ↗
各模型系列，每个一行概览
Anthropic
Claude
Long-context document work, careful reasoning, writing and coding; Projects/Styles carry your profile.
OpenAI
GPT
The broadest all-rounder; Custom Instructions, memory and Custom GPTs make it deeply personalizable.
Google
Gemini
Multimodal strength and Workspace integration; Saved Info and Gems hold your identity.
xAI
Grok
Real-time X grounding and research modes.
Meta (open weights)
Llama
The most deployed open-weight family, self-hostable, powers Meta AI.
Mistral AI (open-weight heritage)
Mistral
Fast European models; agents via La Plateforme.
DeepSeek (open weights)
DeepSeek
Open-weight reasoning models at minimal cost; self-hostable.
"擅长领域"为编辑摘要，并非量化评测。无论哪个产品系列最适合你的工作，你的 Agent of Me 档案始终随你携带。