An AI companion's "memory" is its ability to recall details from earlier conversations. It is the single biggest quality gap between platforms, and it is almost never published as a specification. Without it, every session is a first date.
Two different mechanisms, often conflated
| Mechanism | What it does | Fails as |
|---|---|---|
| Context window | Holds the current conversation | Forgetting what you said an hour ago, mid-scene |
| Persistent memory | Stores facts across sessions | Forgetting your name next week |
Context is a hard technical limit: once a conversation exceeds the window, the earliest content falls out. Persistent memory is a product feature built on top, usually by extracting facts and re-injecting them. They fail differently and platforms rarely distinguish them in marketing.
How to test it in five minutes
Tell the character something specific and arbitrary. Continue for a while. Come back the next day and reference it obliquely. If it survives, persistent memory works. If it survives only within the session, you have context and nothing more. Full method in how to test before paying.
Why cheaper tiers feel worse
Context is expensive to serve, so lower tiers often silently run a smaller window. The model is the same; the memory is not. This is a common and undocumented reason a platform feels worse than the reviews suggested. See how credit pricing works and what an AI girlfriend app is.
Part of the full NSFW AI safety checklist.
FAQ
Why does it forget mid-conversation?
You exceeded the context window and the earliest turns dropped out. Longer roleplay makes this worse, and cheaper tiers hit it sooner.
Can I make it remember something permanently?
Some platforms expose a persona or lore field that is always injected. That is more reliable than hoping a detail survives extraction.
Is more memory always better?
Broadly yes, but it costs credits or tier. The honest question is whether the platform tells you what you are getting, and most do not.