How Does AI Memory Work Across Chats? What Persists and What Doesn't

The thing nobody tells you

You'd expect a single assistant to carry a thread across chats the way a person carries a conversation from room to room. It doesn't. AI memory across chats is asymmetric and confusing precisely because it's not a memory — it's a window that resets, plus whatever a platform happens to bolt on top.

Here's the honest model: within one chat, the assistant sees everything in its current context. Close the chat, and all of that is gone. The only AI memory across chats that actually exists is the stuff stored somewhere the model can find again.

This is worth understanding properly, because it explains a lot of the weird behaviour you'll see — an assistant that remembers a preference one day and forgets it the next, or that seems to know you on one device and not another.

What persists vs. what doesn't

Let's split it cleanly.

Within a single chat: - Recent turns are visible and usable. - Earlier turns survive until the window fills. - Once the window sheds them, they're simply gone.

Across chats (the default): - Almost nothing carries over on its own. - A new chat starts blank unless the platform has a memory feature. - Even when a platform adds a memory feature, it's usually a summary or a short list of facts, not a full record.

So the honest answer is that most "memory" you experience is either the current window or a shallow, platform-controlled scrapbook. Neither is a real memory.

Here's the asymmetry that confuses people: an assistant can be great at recalling within a session and terrible at recalling across one. That's not a contradiction. It's two different storage systems — one temporary and one that barely exists.

Why summaries and fact-lists fall short

The memory features big platforms ship are better than nothing, but they're limited:

• Fact lists lose the shape of a relationship. You can store "user likes X," but you lose the nuance of how and why.
• Summaries are lossy retellings. A snapshot of a conversation is a paraphrase with someone's judgment baked in about what mattered.
• They're cut off or capped. Most have a ceiling, and they don't reliably pull from your whole history.
• They don't distinguish what matters. Everything is stored at roughly the same importance, so the useful stuff gets lost in the noise.

You end up with a scrapbook, not a record. A scrapbook is nice — it's better than nothing. But it's not a memory, and it won't survive the way a real one does.

A concrete example

Say you've been using an assistant for months on a personal project. You've told it your preferences, your constraints, your working style. You've had a long conversation where you worked out the details.

Now you open a new chat and ask it to continue. What does it actually know?

If it has the default setup — probably very little. Maybe a few facts it remembered. It definitely doesn't have the full arc of the project. It's working from a handful of notes, not from a relationship.

The frustrating part is that it feels like it should remember, because it's the same assistant, same login, same account. But the account is just a handle. The memory is whatever got saved, and most of the time, very little got saved.

What a real cross-chat memory looks like

A genuine persistent memory is different in kind, not degree:

• It's a store, not a summary. The meaningful source is retained, not compressed into a highlight reel.
• It's retrievable on demand. Relevant history gets pulled into a fresh window when it's useful, instead of all being held at once.
• It survives indefinitely. It's not capped at a few facts; it keeps what you decide matters.
• It distinguishes importance. It can hold both the big picture and the small details, and surface the right one at the right time.
• It's yours. Ideally local, so the record is private and under your control.

That's how an AI goes from "chatty" to "actually knows you."

How to tell whether you have it

The simplest test: tell the assistant something specific and personal in one session. Then, in a fresh session days later, ask whether it knows that thing. If it does, there's a real store behind it. If it doesn't, you've been relying on a window — and you've been doing more of the remembering than you realised.

The bottom line

AI memory across chats doesn't happen by default. The window resets, and the only carry-over is whatever a platform chooses to keep in a shallow, capped store. The reliable answer is a persistent memory layer — a real record outside the window.

Read next: the architecture of remembering between sessions and what real AI persistent memory takes.