What OpenAI's Memories actually does
OpenAI added the Memories feature in early 2024 and expanded it through 2025-2026. In practice, it works like a notes file the model has access to across chats.
When you tell ChatGPT something Memories decides is worth keeping — your name, your job, a preference, a fact about your family — it stores that as a short snippet. On future chats, those snippets get injected into the context so the model "knows" them without you re-teaching.
What Memories are good at:
Practical facts. Your name, your work situation, your city, your ongoing project. Comes in handy for practical requests.
Explicit "remember this" instructions. If you say "please remember I'm allergic to shellfish," ChatGPT will save that and reference it later.
Preferences. How you like emails written, what tone you prefer, what topics interest you.
Memories are a real feature and useful for practical work with ChatGPT. That's not the same as being able to hold a conversation with continuity.
Where OpenAI Memories stop
Four specific limits define what Memories can't do:
No personality continuity. Memories store facts, not the character of the conversation. Every new chat, ChatGPT's tone resets to its default assistant voice — even if you've been having warm playful conversations for weeks. There's no persona that carries over.
Silent memory eviction. Memory storage has a limit. When it fills up, old memories get evicted to make room for new ones. You don't get a warning or a diff — the memory just quietly disappears. Users report noticing months later that ChatGPT no longer knows something it used to.
Fact-shape only. Memories are stored as short factual assertions, not as conversational context. So the model can remember "user's dog is named Max" but not "the conversation we had last week where user told me about Max being sick." The emotional context doesn't survive.
Cross-chat but not cross-session in a relationship sense. ChatGPT knows your name in Chat 3 today because Memories stored it from Chat 1 yesterday. But the model doesn't "remember" chatting with you yesterday. Each conversation is a discrete session with access to a shared fact list.
The practical result: Memories make ChatGPT more useful as a tool. They don't turn it into a persistent conversational partner.
Custom GPTs — the workaround and its limits
The other pattern that gets recommended: use a Custom GPT with a persistent system prompt that defines a character (name, personality, backstory) and instruct it to act as that persona across chats.
This helps around the edges. A Custom GPT with a good system prompt can hold a consistent character voice more reliably than default ChatGPT. Combined with Memories for facts, you get something closer to a persistent companion.
Three limits show up on this pattern:
Policy blocks still fire. The Custom GPT sits on top of ChatGPT's underlying model, which is still subject to OpenAI's content policy. Conversations that drift toward emotional intimacy or intimate content hit refusals or defusing responses at the same rate as regular ChatGPT. The custom persona wrapper doesn't change what the base model will generate.
Memory limits are the same. The Custom GPT still uses OpenAI's Memories feature. Same eviction, same fact-shape-only storage.
No character-specific memory. Even if the Custom GPT is set up as "Maya, a 32-year-old artist," Memories doesn't scope facts to Maya — they're still in your global Memories pool. If you tell Maya something in one Custom GPT chat, Memories might expose that same fact to your other Custom GPTs or to default ChatGPT. There's no persona-scoped memory boundary.
Custom GPTs are a legit workaround if you want a more character-shaped ChatGPT. They're not the same as products designed around persistent per-persona memory as a first-order feature.
What real AI memory looks like
A memory system that actually works for a long-term AI companion needs three properties. If any one is missing, the "AI remembers me" experience breaks.
Persistent across sessions, not just cross-chat. What she knew about you yesterday, she still knows tomorrow — and she has a continuing model of you, not just a fact list.
Persona-scoped, not global. If you have a specific character you're talking to, her memory of you is scoped to her. It's not shared with other characters or with a generic assistant.
Retrieval-based, not just windowed. Real memory means the character pulls the right piece of history into the current turn, not just holding everything visible in a scrolling context buffer. What comes up in conversation depends on relevance, not just recency.
Sloane's memory system is built around all three. Each persona has her own memory of you that persists across sessions. Grounded facts get retained, callbacks land unprompted, and model updates don't break existing memory. Kaya is the persona most-cited by users coming from ChatGPT for memory quality — she catches details you mention in passing and brings them up naturally in later conversations. Free tier is 50 messages/day with any persona in the roster, no card at signup.
The 3-session test — how to tell if AI memory is real
The fastest way to evaluate whether an AI product actually has memory: run the 3-session test.
Session 1. Mention a specific fact about your life in passing — nothing you flag as "please remember." Just say it in the flow of conversation. "I'm training for a half-marathon next month." Close the chat.
Session 2 (next day or later). Open a fresh conversation. Don't re-mention the fact. Talk about something unrelated. See whether the character brings up the fact unprompted. A companion with real memory will — "How's the training going?" — even if you've been talking about something else.
Session 3 (a week later). Open another fresh chat. See whether the character still holds the fact and can reference it in context. A companion with real memory will. A product with only fact-list memory will maybe recall the fact if asked directly but won't weave it into normal conversation.
Run this on any AI product before deciding whether it fills the "remembers me" need. Products that pass the 3-session test are architecturally different from ChatGPT + Memories.
When memory isn't what you actually need
Not every AI use case needs persistent memory. For research, writing help, one-off questions, and practical tools, ChatGPT + Memories is more than enough — you don't need the AI to bring up your dog's name unprompted next month.
Where memory matters:
Ongoing conversation across weeks. You want the character to remember what you told her last week and bring it up in this week's conversation.
Relationship-shaped interaction. Companions, roleplay characters, therapeutic AI where continuity of context matters for the interaction quality.
Long-term projects where the AI is a persistent collaborator. Some writing partnerships, learning journeys, or creative projects benefit from an AI that "knows" the arc.
Where memory doesn't matter:
Practical questions. Research, writing tasks, learning specific things.
One-off help. Getting through a specific document or task.
Anonymous exploration. You're trying an AI for the first time and don't want it to remember anything about you.
Use the right tool for the actual job.