The AI Companion Memory Test: 5 Prompts That Expose Fake Memory

Every AI companion app claims to remember you. Instead of taking that on faith, here's a concrete, repeatable test you can run on any app in about ten minutes across two sessions.

1. Plant a specific, unusual detail

Early in a conversation, mention something oddly specific rather than generic — not "I like coffee" but "I only drink coffee if it's from a blue mug, otherwise it tastes wrong to me." Specificity is what a summary-based system tends to lose first.

2. Wait — don't test it in the same session

Anything can seem to "remember" something you said five messages ago; that's just the context window doing its job. The real test starts after the session ends and enough time passes that the original conversation is no longer sitting in active context — a day is enough, a week is more convincing.

3. Ask an indirect question, not a direct one

Instead of "what mug do I drink coffee from?" (which cues the AI on exactly what to retrieve), ask something adjacent, like what the character would get you if they were making you a coffee. Real retrieval surfaces the detail unprompted; a system guessing from a vague summary tends to answer generically or confidently make something up instead.

4. Check whether it holds up to gentle pressure

If the character recalls the detail, follow up and ask when you mentioned it or what else you said around the same time. Systems papering over a lost memory with a plausible guess tend to fall apart under a second question; systems retrieving an actual stored fact tend to stay consistent.

5. Repeat it a second time, weeks later

A single successful recall could be luck or a broad category match. Repeating the test with a different specific detail, and checking it again further out, is what actually tells you whether an app has real persistent memory or just a memory system that works well over short gaps.

What this test is really checking

All five prompts are checking the same thing: whether the app stores structured, retrievable memory that survives outside the current context window, or whether it's reconstructing a plausible-sounding version of "remembering" from a recent summary. Vantrix characters are built on the former — memory as its own retrievable system, not a byproduct of a longer context window — which is exactly why this test is one we're comfortable pointing people toward.