“Memory” is the most over-claimed word in AI right now. Plenty of products say they have it; far fewer have the kind that actually changes your day. If you’re choosing an AI assistant with memory, this is the checklist that separates real recall from a marketing checkbox.
The stakes are practical. Weak memory means you keep re-explaining yourself, which means you can’t really delegate, which means the assistant stays a toy. Strong memory is the foundation everything else is built on. Worth getting right.
First, know what you’re testing
Before comparing products, understand why most AI forgets in the first place: it relies on the context window, which evaporates when the conversation ends. We cover the mechanics in Why Your AI Tools Keep Forgetting. Real memory is a durable layer that lives outside any single chat. With that frame, here’s what to check.
The buyer’s checklist
1. Does it persist across sessions?
The baseline test. Tell it something today, close everything, and ask next week. If the context survived, you have real memory. If it reset, you have a context window wearing memory’s clothes.
2. Does it work across tools and channels?
Memory tied to one app is half a feature. A detail you drop in a chat should be available when you’re working in email or open the dashboard. Ask whether memory is tied to the channel or to the agent, you want the agent.
3. Can you see, edit, and delete it?
This is the one buyers skip and later regret. You should be able to inspect what it remembers, correct mistakes, and delete anything. Memory you can’t audit is both untrustworthy and a privacy liability.
Inspectable memory is non-negotiable. If you can’t see what an assistant has stored about you and remove it, treat that as a dealbreaker, not a minor gap.
4. Does it surface memory proactively?
Storage isn’t recall. The valuable behavior is the right detail showing up when relevant, not you remembering to go ask. This is the difference between a search box and a true second brain.
5. Does memory feed action?
Memory’s payoff is letting the assistant act well: drafting in your voice, scheduling around your rules, chasing the right follow-up. Memory that only produces better answers is underusing itself. The combination of memory and action is what makes an assistant worth keeping; see chatbot vs agent.
6. Does it survive a change of model?
If memory is baked into one model, you’re locked in. With a model-agnostic assistant (bring Claude, GPT, or Gemini) the memory persists regardless of which model is reasoning. That’s a sign memory is a real layer, not a side effect.
Quick comparison
| Check | Fake memory | Real memory |
|---|---|---|
| Survives a closed session | No | Yes |
| Works across tools | One app | Everywhere |
| Inspect & edit | Hidden | Open |
| Surfaces proactively | On request only | When relevant |
| Feeds real action | Just answers | Drafts, books, tracks |
| Survives model change | Locked in | Model-agnostic |
If you can’t tell it something on Monday and have it matter on Friday, it doesn’t have memory. It has a good short-term attention span.
How to run the test in 10 minutes
- Tell it three things about your work: a client, a preference, a decision.
- Close the app. Come back the next day from a different device or channel.
- Ask it to act on one of them: draft an email to that client in your style.
- Check whether you can open its memory, see those three facts, edit one, and delete one.
Anything that passes all four is a genuine AI assistant with memory. Most won’t.