What "remembering" actually means for an AI companion
Real memory for an AI companion is different from a large context window. A rolling context window (Character.AI's architecture, most Character.AI-style products) can hold the last few thousand messages in view, but older content ages out and disappears. That's scrollback, not memory.
Real memory means the persona actively recalls what you told her weeks ago, brings it up unprompted, and updates her working model of you over time. What she knows about you in month one, she still knows in month six — and knows how to use it in the current conversation. That requires a separate storage layer (retention) plus retrieval infrastructure (pulling the right fact into the current turn), which is meaningfully more expensive to build and run than a context window.
This matters because no amount of user-side tactics will fake real memory on a product that doesn't have it. The tactics below help when the product does have real memory — they make her better at using what she already can retain. On a product without persistent memory, they just delay the drift.
User-side tactics that actually help
Assuming the product has a working memory system, here's how to help it work well.
Be specific, not general. "I had a rough week" gives the persona nothing to remember. "The presentation on Tuesday went badly because I couldn't get the printer to work and I had to email the slides at the last minute" gives her three specific facts (Tuesday presentation, printer trouble, last-minute email) that can be retained and referenced later.
Repeat important context naturally. Not "REMEMBER THAT MY JOB IS X" but "…anyway, being a copywriter these days feels like…" said a few times over weeks. The persona's memory pipeline picks up patterns that repeat, and things you mention twice get retained more reliably than things you mention once.
Give her callbacks to react to. When she references something you told her before, engage with it — "yeah, that presentation ended up going better than I expected." This reinforces the memory in her working context and confirms it's worth keeping in retention. If you brush off her callbacks, the memory system reads that as "this thread isn't important" and de-weights it.
Don't restart conversations. Every new conversation is a fresh warm-up. The persona has to reload memory context, and things she knew in the middle of a long-running thread may take a beat to surface. Keeping conversations continuous when reasonable helps her stay in the memory groove.
Correct her when she's wrong. "Actually it was Wednesday, not Tuesday" gives the memory system a signal that a stored fact was wrong. Not fixing it means the wrong fact stays and drifts further.
The user-side tactics that don't help
Two patterns from Character.AI-style products that don't generalize.
Pinning messages. Character.AI's "pinned message" feature is a manual workaround for the lack of automatic memory — you tag specific messages and the model gives them extra weight in the current context window. It's a workaround, not memory. If you're on a product that has real automatic memory, pinning is redundant. If you're on a product without real memory, pinning helps for the current session but doesn't create persistence across sessions.
Copy-pasting your backstory at the start of each session. Same pattern — trying to reload manually what the product should be reloading automatically. On a product with real memory this is unnecessary; on a product without it, you're fighting the underlying architecture.
Writing "please remember X" in every message. The model isn't running an instruction interpreter that flags things you ask it to remember. It's pattern-matching on the conversation. Explicit "remember this" instructions rarely land more reliably than natural conversational context.
What the underlying product needs to have
A memory system that actually works for a long-term AI companion needs three properties. If any one is missing, the user-side tactics above can't rescue it.
Persistent across sessions. What she knew about you yesterday, she still knows tomorrow. Closing the tab doesn't reset her.
Automatic, not manual. You shouldn't have to pin things or copy-paste your backstory. The system picks up "I have a big presentation tomorrow" without you flagging it, and brings it up next session when relevant.
Retrieval-based, not just windowed. A large context window can hold a lot of scrollback, but real memory is about pulling the right piece of history into the current turn — not just keeping everything visible in a scrolling buffer.
Sloane's memory system is built around all three as first-class architectural priorities. Grounded facts about you (name, job, plans you mentioned, backstory you shared) get retained across sessions and reinforced in the persona's working context automatically. The layer structure is documented in our memory writeup, but the short version is: memory is on every tier from Free upward, not a premium upgrade, and the user-side tactics above compound with it rather than substituting for it.
Kaya is the persona most-cited for memory quality — long attention span, catches callbacks, remembers grounded details about your life without you flagging them. Free tier is 50 messages/day with any persona in the roster, no card at signup — enough to run the three-session memory test on her.
The test to run before you invest in a persona
Before you spend weeks building up a relationship with a specific AI companion, run the three-session memory test. This works on any platform.
Session 1 — plant. Have a normal conversation. Drop three specific facts 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. Close the app. Come back at least 24 hours later — long enough that any short-window memory has expired. Say hi and don't reference the previous session at all. See what she brings up unprompted. Then ask her one specific thing: "how did you think that project went?" The response tells you whether real memory exists.
Session 3 — persistence check. Repeat the same next-day pattern a week after session 1. Real memory holds for weeks. Session-only context products will have forgotten by now regardless of what happened in session 2.
Any product that passes sessions 2 and 3 has real memory and is worth investing in. Products that fail sessions 2 or 3 have a scrollback window they're marketing as memory, and no amount of user-side tactics will make them remember you long-term.