Which AI Models Actually Have Memory? The Honest List

The short answer

Which AI models have memory? None of them — and that's the first thing to understand, because it saves you from a lifetime of wrong expectations. Every major AI model is stateless: it holds nothing between conversations. What people actually mean when they ask "which models have memory" is which products layer memory on top of a model — and that's a very different list, with very different answers.

Here's the honest version of that list, what each product actually keeps, and the three questions to ask before you trust any of them.

The myth first: models don't remember, products do

A model is a function. You give it input; it produces output; it keeps nothing. The "context length" you see on a model card — 128K, 200K, a million tokens — is the size of the working view it can hold in one session, not a memory of you. A bigger context length means a bigger whiteboard, and every whiteboard gets erased. That distinction is the whole ballgame, and it's explained properly in context window vs context length.

So when a company says "our model has memory," translate it: they've built a store, a write pipeline, and a retrieval system around a stateless model, and they're calling the whole package "memory." The model isn't remembering. The system is. Which means the honest question isn't which model — it's which product has built memory well.

The honest list: what popular AI products actually keep

Here's the plain-spoken state of play. Features change fast, so treat this as a map of the territory as of September 2026 — and verify the current details before you bet anything important on them.

Product	What it actually keeps	What it loses	Verdict
ChatGPT	A list of facts it decides (or you tell it) matter; Projects hold instructions and files	The full thread of any conversation	Useful facts, shallow continuity
Claude	Memory facts; Projects with pinned instructions and uploads	The reasoning and decisions inside a chat	Solid project container
Gemini	Memory of stated preferences and facts in the consumer app	Conversation histories beyond what the feature stores	Same fact-list pattern
Cursor	An indexed view of your codebase; rules files you write	Decisions made in chat sessions	Great code context, no chat memory
Claude Code / coding agents	Whatever you put in CLAUDE.md / AGENTS.md and state files	The session itself	As good as your discipline
Companion apps	Varies wildly — some keep real episodic records, most keep a thin fact profile	Depends entirely on the architecture	Read the privacy policy

The pattern should jump out: no row in that table remembers your conversation by default. The products that come closest are the ones where you build the container — Projects, rules files, state files — or where the vendor has invested in a genuine persistence layer rather than a fact list. The difference between "remembers facts about you" and "remembers you" is the difference between a post-it note and a diary, and it's the core of the types of agent memory.

How to read the marketing

Every product in this space says "memory" somewhere on its pricing page. Here's the decoder ring:

• "Remembers your preferences" → stores a semantic fact list. Shallow, but honest.
• "Remembers your conversations" → claims episodic memory. Ask what it actually retains — full source, or summaries? (Summaries drift. The continuation problem explains why.)
• "Context window of a million tokens" → not memory at all. It's a bigger whiteboard, and it still gets erased.
• "Works across your codebase" → retrieval, not memory. Great for code, useless for remembering yesterday's decisions.
• "A model that remembers" → marketing. No model does. Ask about the system around it.

The most expensive mistake is choosing a model for its "memory" and discovering later that you meant persistence all along. Memory is a system property, not a model property — which is exactly why this site keeps coming back to the persistence thesis.

The three questions that matter more than any list

Forget which model "has memory." Ask these three things about any AI you're considering trusting with continuity:

1. What does it store — facts, history, or both? A fact list ("I work in marketing") is not a history ("we decided to reposition the brand in March"). If your work lives in decisions and context, you need a product that keeps more than facts.

2. Where does the memory live — and who can read it? Vendor memory lives on vendor servers, subject to vendor policy. If the content is sensitive, that's a real question, and the answer might be local AI agent memory or another local-first option.

3. Can you read, edit, and export the record? A memory you can't inspect is a memory you don't own. The best designs are transparent — you can see exactly what's been saved about you, correct it, and take it with you if you leave. Anything else is a loyalty program dressed as a brain.

What actually has memory

If you want the real answer to "which AI has memory," it's the one you build around. Take any capable model and give it a durable store — files, a database, a persistence layer — and that system has memory, regardless of the model underneath. This is why the serious conversation about AI continuity isn't about model releases at all; it's about the architecture around them. See how does AI memory work for the mechanism, and the memory layer for AI agents for the component that makes it real.

The practical upshot: choose your model for intelligence, your product for its persistence design, and your own habits for everything in between. The best "memory" in the world is a capable model plus a store you control.

The bottom line

Which AI models have memory? None — models are stateless by design. Which products layer memory on top? ChatGPT, Claude, Gemini, Cursor, and the coding agents all do, in different shapes and depths: fact lists, project containers, code indexes, and rules files. Read the claims with the decoder ring, ask the three questions, and remember that real memory is architecture, not a model feature. Build your own layer around a good model, and you'll have the only answer that matters.

Read next: how does AI memory work, and the types of agent memory.