Memory Guide
ChatGPT Memory — Complete Guide (2026) logo

ChatGPT Memory 完全指南(2026)

2026 年 ChatGPT Memory 工作原理全解:存什么、忘什么、如何控制,以及什么时候应该自建记忆层。

为什么选它

ChatGPT memory is a closed feature inside chatgpt.com and the ChatGPT mobile apps. It is not a developer API. When you read "ChatGPT remembered my name across chats", you’re looking at two different mechanisms: saved memories (explicit facts ChatGPT wrote into a list you can review in Settings) and chat history reference (a newer feature that lets the model retrieve from your past conversations).

Saved memories are short, declarative, user-scoped: "User prefers Python over JavaScript", "User has a daughter named Ana". The list is editable — Settings → Personalization → Memory. You can delete individual entries, turn memory off entirely, or ask the bot directly to "forget that".

For developers, ChatGPT memory does not exist: the OpenAI API has no memory endpoint. If you’re building your own product and want ChatGPT-like memory, use mem0, Zep, or Letta. This guide is for consumer users of ChatGPT, plus developers who want to understand what OpenAI built so they can replicate the UX.

How to control ChatGPT memory (consumer)

There’s no API for ChatGPT memory — these are UI actions only. Memories are stored in OpenAI’s infrastructure and tied to your ChatGPT account; they don’t sync with the OpenAI API, Copilot, or any third-party app.

# In chatgpt.com / mobile app
# Settings → Personalization → Memory

# Check what's remembered
"Show me what you remember about me."
→ ChatGPT lists every saved memory entry.

# Add a memory explicitly
"Remember that I prefer TypeScript for frontend and Go for backends."
→ Saves 1-2 entries in the memory list.

# Forget a specific memory
"Forget what you remember about my current project."
→ Removes matching entries; it may ask for confirmation.

# Full reset
Settings → Personalization → Memory → "Clear ChatGPT's memory"
→ Wipes all saved memories. Chat history reference (if on) remains.

# Opt out entirely
Settings → Personalization → Memory → toggle off
→ No new memories saved; existing ones stop being used.

# Temporary chat (incognito mode)
Click temporary chat in the chat sidebar
→ That conversation neither reads nor writes to memory.

核心能力

Saved memories (Settings-visible)

Short textual facts the model chose to save. Fully visible and editable. Typical size: a few dozen entries, each 1-2 sentences long.

Chat history reference

Separate from saved memories. Allows the model to retrieve relevant past chats on demand. Toggle in Personalization. Rolling out to all paid users in 2025-2026.

Per-GPT memory isolation

Custom GPTs have their own memory scope — a memory saved in "Product Strategy GPT" won’t leak to the default ChatGPT. Useful for keeping work and personal contexts separate.

Temporary chats

One-shot conversations that neither read nor write memory. Equivalent to an incognito session. Use when discussing sensitive topics you don’t want remembered.

Privacy controls

Opt out entirely, delete individual entries, or clear all memories. OpenAI’s data policy governs retention after deletion; see the ChatGPT help center for current terms.

Works across devices

Memories tied to your OpenAI account sync across web, iOS, Android, and desktop clients. Not available in the API, Playground, or third-party wrappers.

对比

 AudienceAPI AccessEditable?Best for
ChatGPT MemorythisConsumer users of chatgpt.comNoYes (UI)Personal assistant UX
mem0DevelopersYes (OSS + SaaS)Yes (API)Your own product
ZepDevelopersYes (OSS + SaaS)Yes (API + UI)Production chat apps
Claude ProjectsConsumer users of claude.aiNoVia Project contextWorkspace-scoped memory

实际用例

01. Personalizing your own ChatGPT

Explicitly tell ChatGPT what to remember — preferences, writing style, ongoing projects. Review the memory list monthly and prune stale entries.

02. Separating work contexts

Create one Custom GPT per domain (e.g., "Fitness Coach", "Code Reviewer"). Each has isolated memory, so preferences don’t bleed between contexts.

03. Understanding what to build

Developers studying ChatGPT’s UX to replicate it in their own product — features like "show me what you remember" and "forget X" are simple patterns worth copying.

价格与许可

ChatGPT memory: included with ChatGPT Free (limited), ChatGPT Plus, Team, and Enterprise. Not available through the OpenAI API. Pricing is the same as your existing ChatGPT subscription.

Enterprise data handling: Enterprise/Team plans include admin controls for memory (org-level toggle, audit export). See the OpenAI trust portal for data residency and retention details.

Developer alternative cost: to replicate ChatGPT memory in your own app using mem0 or Zep, expect $0.0002-$0.001 per user turn in LLM extraction cost + small storage fees. See the respective pricing pages.

相关 TokRepo 资产

常见问题

Can I access ChatGPT memory from the API?+

No. The OpenAI API (chat completions, assistants, responses) has no access to ChatGPT memory. The memory feature is exclusive to chatgpt.com and the ChatGPT apps. To add memory to your own app, use mem0, Zep, Letta, or LlamaIndex.

How much does ChatGPT remember?+

Saved memory holds dozens of short entries, not unlimited history. Chat history reference can pull from any of your past conversations but is retrieval-based, not always-on context. Plus/Team have higher limits than Free.

Does ChatGPT memory work with plugins / Custom GPTs?+

Custom GPTs have isolated memory scopes. Plugins (now called "actions") can read user messages but do not directly access memory entries — they only see what ChatGPT puts in the prompt.

How do I opt out of memory entirely?+

Settings → Personalization → Memory → toggle "Memory" off. Also disable "Reference chat history". In Enterprise/Team accounts, admins can enforce this at the workspace level.

Is ChatGPT memory safe for sensitive info?+

For consumer Free/Plus, assume memory is logged on OpenAI infrastructure per OpenAI’s data policy. For genuinely sensitive data (health, financial, legal), use temporary chats, or pay for Enterprise/Team plans with stronger data guarantees. Developers building their own apps keep data inside their own infrastructure when they use self-hosted memory layers.

Why would I build my own memory when ChatGPT has one?+

Because ChatGPT memory lives in OpenAI’s product. If you’re building your own AI app, you need memory in your own product — with your own users, branding, data model, and controls. Every developer-facing memory library exists because "use ChatGPT memory" was never an option.

同类推荐