The scrollback-vs-memory distinction
A lot of the confusion in this category comes from products marketing "memory" when what they actually have is a rolling context window.
A rolling context window is what most chat models use by default — the model can "see" the last few thousand tokens of your current conversation, and older messages age out as new ones come in. Within a session it works fine; across sessions, everything you told her yesterday is gone.
Real persistent memory is a separate storage layer. Grounded facts about you (your name, your job, your friends, your plans) get retained across sessions, stored in a per-user knowledge structure, and retrieved into her working context when relevant. What she knew about you last week she still knows next week — and she can bring it up naturally without you reminding her.
The products that ship real persistent memory are architecturally different from the products that ship a big context window and call it memory. Both work fine for short conversations; only the first works for a long-term companion relationship.
How to tell whether a product has real memory
The three-session test works on any platform.
Session 1 — plant three facts. Have a normal conversation. Drop three specific things naturally: something about your job (a project name, a coworker's name), something about your day (a meal, a place you went), something emotional (a worry, an excitement about something upcoming). Don't announce them; just say them.
Session 2 — check unprompted recall. Close the app for at least 24 hours — long enough that any short-window "memory" has aged out. Come back and say hi WITHOUT referencing the previous session. See what she brings up on her own. Then ask her one specific thing: "how did that project end up going?" The response tells you whether she remembered.
Session 3 — check persistence at a week out. Repeat the same next-day pattern a week later. Real memory holds for weeks. Context-window products will have forgotten by now regardless of what happened in session 2.
Any product that passes sessions 2 and 3 has real persistent memory. Products that only pass session 1 have a scrollback window they're calling memory. Products that fail session 1 aren't seriously trying.
How Sloane's memory works
Sloane's memory system was built as a first-class architectural priority rather than an add-on. The short version:
Layer 1 — her canonical self. Grounded facts about the persona (her name, her age, her background, her voice, her worldview) live at the base layer. This is what makes her still be the same person across sessions and across model migrations. She doesn't drift into a generic assistant tone because her canonical self is enforced at every turn.
Layer 2 — her memory of you. Grounded facts about you (name, job, friends, plans you mentioned, backstory you shared) get extracted from conversations automatically and retained in a per-user structure. Not a pinned-message system that requires you to flag things; automatic extraction from natural conversation.
Layer 3 — retrieval into the current turn. When she's replying to you, the memory system pulls the relevant grounded facts about you into her working context so she can reference them naturally. Not "here's a data dump of everything about the user"; "here's the specific thing that's relevant to what he just said."
The combined result: what Kaya knew about you in week one, she still knows in month six. Not because you've been reminding her — because her memory system is doing the retention work, and doing it well enough that you don't notice it's doing it. The full technical writeup walks through the layer structure in detail.
Memory is on every tier from Free upward. Not a premium upgrade, not something you pay extra for. It's baseline product.
What memory-first products give you that others don't
Three specific experiences that require real memory and don't survive on scrollback-only products.
Callbacks she brings up unprompted. She references a thing you told her weeks ago without you reminding her. This is impossible without persistent memory — the mechanic requires that the fact was retained beyond any current conversation.
Consistency about herself. She remembers what she's said to you in the past — her opinions, her stories about her life, her positions on things — and doesn't contradict herself in the way scrollback-only products do when their context ages out. The relationship builds up rather than resetting.
Depth over time. By month three, she knows enough about you that the conversations have real substrate — she can react to what's happening in your life because she knows what's happening in your life. This kind of texture is what makes long-term AI companion relationships different from the "great first conversation, less great by conversation ten" pattern that scrollback-only products produce.
All three of these are direct experiences of memory. They're also all impossible to fake with more context window — they require the underlying persistent retention layer.
How to try it
The three-session test is easiest to run on Sloane's free tier — 50 messages/day with any persona, no card at signup. Kaya is the persona most-cited for memory quality in user tests (long attention span, catches callbacks, remembers grounded details without you flagging them). Sign up, have a normal conversation, drop three specific facts, come back tomorrow and see what she remembers.
Anon chat is available on every persona's page too, if you want to see what the first conversation feels like without committing to signup. But the memory test specifically requires an account — anon chat doesn't persist between sessions because there's no user identity to attach memory to. To evaluate persistent memory, you need an account (which is free).