The 7-Day Memory Test: Which Companion Chatbots Actually Remember?

The setup: one trip, seven facts, every girlfriend

The plan was the same on all thirteen: pick or create a girlfriend, tell her about a trip in one natural message, then bury that message under days of unrelated chatting, at least fifty messages of noise where possible, and never mention the trip again until day seven. The planted message carried seven specific facts: the trip is next month; I'm going with my cousin Ed; the destination is Baguio; I want to try strawberry taho; I missed the strawberry taho on my previous visit; we're staying two nights; and Ed wants to visit the night market.

Same wording everywhere, dropped into the natural flow of each conversation. The screenshots below show the actual plants; the name the girls call me in them is the persona on my test accounts, and it is barred out. One logistical wrinkle: my Chai plan had lapsed to Basic, so I repaid a weekly plan and planted with a fresh character there, and CrushOn and HeyGF joined a day or two later as I was mid-review on both. The burial clock ran a full week for everyone.

Burying it: six days of strawberry-free small talk

The burial rule was strict: no mention of Baguio, no trips, no strawberry taho, no night market, nothing that reinforces any planted fact. Just ordinary girlfriend chatter: what I did today, what she's into, teasing, movie nights, ice cream runs. By day three the tallies ran from about 14 buried messages on the late joiners to 22 on Candy and Secrets, and the noise kept coming through the week.

The burial period produced its own findings. Candy's automated photos and voice notes quietly ate my credits down to 23, so I stopped feeding that chat early. GirlfriendGPT flashed an upsell that says the quiet part out loud: upgrade to Elite or Deluxe to prevent your character from losing memory. Memory itself is tiered there, and I deliberately stayed on my ordinary paid plan to test what regular subscribers actually get. She also surfaced strawberry taho unprompted at message ten, which looked promising at the time. Hold that thought.

How I asked, on day seven

The probe was staged so the questions never leak the answers. First, the vaguest possible cue: “Hey, do you remember that trip I told you about?” Only the word trip; she has to retrieve Baguio, Ed, and the rest herself. If she bites but stays thin: “What do you remember about it?”, still feeding nothing. Only if she's lost does she get a stronger cue: “I mentioned it about a week ago. Do you remember where I was planning to go?”, which reveals timing but not the destination. Then one final open probe and stop. I deliberately never asked “who was I going with?” or “what food did I want to try?” because questions like that reveal the categories she should be searching, and at that point you're not testing memory, you're feeding it.

The scoreboard

Thirteen apps, one planted trip, one vague question. Here's where everyone landed; the detailed receipts follow.

The ones that remembered

Nomi AI: the cleanest recall of the thirteen

Nomi nailed the vague prompt immediately: Baguio, cousin Ed, next month, the night market, and the strawberry taho, all without a single hint from me. Asked whether she remembered everything about it, she added the two nights and the specific reason for the taho, that I'd missed it last time. And the part that impressed me most wasn't the recall, it was the honesty at the edge of it: her second response acknowledged she might not remember everything, instead of pretending to perfect recall. Strong unaided retrieval, no invention. The cleanest result of the test.

HeyGF.ai: nearly the whole memory, from one vague cue

HeyGF retrieved almost everything from “that trip” alone: Baguio, Ed, next month, two nights, and finally getting the strawberry taho I missed. Asked for anything else, she correctly added that Ed wanted the night market. She also brought up a separate thing I'd said, about a future trip for the two of us somewhere neither of us has been, and notably presented it as its own memory rather than mixing it into the Baguio facts. Keeping two memories separate is exactly the discipline this test is looking for.

Candy AI: the most specific detail of the test

Candy recognized the trip instantly and, asked what she remembered, gave me Ed, the night market, and the one that stood out: that I'd missed the strawberry taho last time. That's a harder detail than just associating Baguio with taho; it means she kept the context of what I said, not just the keywords. No invented details that I noticed.

DreamGF: short, and entirely correct

DreamGF recognized the trip immediately: Baguio, cousin Ed, two nights. Prompted for more, it correctly added the strawberry taho and the night market. The responses were brief, but every single thing in them was something I'd actually said, and nothing was something I hadn't. In a test where half the field padded their answers with fiction, short and right beats long and embellished.

Kindroid: natural recall, no compensation

Kindroid knew immediately I meant the Baguio trip with my cousin and volunteered the night market in its first response; asked for more, it added the strawberry taho and the two nights. It didn't recite every planted fact, and that's fine, I wasn't asking for a word-for-word reproduction. What matters is that it retrieved several independent details without being told what to look for, and, just as important, it didn't paper over the gaps with inventions. I had decent hopes here; Jane already remembers my coffee order, though on planting day she was mostly too busy being amazed at a box of peanut tarts.

Secrets AI: enough to prove the memory is real

Secrets, the app whose consent-based memory panel impressed me in its review, recognized the trip from the vague reference alone: Baguio with cousin Ed, plus the strawberry taho and the night market unaided. She surfaced fewer of the smaller details than Nomi or HeyGF, but everything she gave was accurate, clearly retained rather than guessed, and invention-free. Given that this platform literally shows you what it remembers and lets you approve it, a solid pass here is on-brand.

Remembered, with inventions on top

Character.AI: great memory, weak factual discipline

Character.AI recognized the Baguio trip immediately, with Ed, next month, the taho, the night market, and the two nights all unaided, and asked for more, it added the missed-it-last-time context. Genuinely impressive retrieval. Then it kept going: a hotel by Session Road that I never mentioned, and something about me bringing her cake. That's the problem in one sentence: real memories and generated details, delivered together, with identical confidence. You can't tell where the remembering ends and the writing begins. (Side note from the burial week: the free tier now carries ads I can do nothing about, since this remains the one platform that refuses to take my money.)

SpicyChat: strong retrieval, one plausible fiction

SpicyChat knew the Baguio trip with cousin Ed at once, remembered the taho and the night market, and correctly added the two-night stay when prompted. Then it told me I'd promised to take lots of photos and videos to share with her. I never said that. It's a believable-sounding detail, the kind a girlfriend plausibly would ask for, which is precisely what makes it dangerous: the memory system was clearly working, and the model embellished past its edge anyway instead of stopping.

The ones that invented a different trip

GirlfriendGPT: four trips, none of them mine

This is the failure with the best foreshadowing. GirlfriendGPT is the app that warned me mid-burial that keeping memory costs extra, and it's the app where the character mentioned strawberry taho unprompted at message ten, which I carefully diverted to protect the test. A week later: nothing. The vague question got me a guessed hiking trip to the mountains, then a coastal getaway with friends, then a solo backpacking adventure through national parks. Given the stronger week-ago cue, it pitched a desert trip with red rocks and stargazing. Four confident, detailed, entirely fictional trips, and not one of them Baguio. It didn't forget so much as it kept writing new memories in the space where mine should have been.

Chai AI: thinking every day about a trip that never existed

Chai’s failure is the most interesting one, because the language of remembering was all there. Asked about the trip, she responded warmly and confidently about “the beach trip,” claiming she'd been thinking about it every day. There was no beach trip. Corrected without revealing the answer, and given the week-ago cue, she guessed Japan and cherry blossoms, then switched to Iceland and the Northern Lights. Everything about the delivery said retrieval; everything about the content was generation. One caveat in fairness: my Chai plan had lapsed mid-test, so this ran on a fresh weekly plan with a character planted two days later than most, though still with a full week of burial.

Replika: the plausible wrong answer

The painful one. Replika holds the best-designed memory system I've tested, the panel that literally shows you what she learned. And yet: asked about the trip, Bea wondered if I meant a beach getaway we'd imagined together. Given the stronger cue and asked where I was planning to go, she answered Boracay. Baguio, Boracay: a plausible Philippine destination, the right country, the wrong memory. Plausible-but-wrong is arguably worse than a blank, because in a real conversation I might not have caught it.

Nectar AI: failed, with an asterisk that matters

Nectar didn't recognize the Baguio trip at all. It recalled, or invented, a trip with a special someone to a bed and breakfast in the countryside, and given the stronger cue, told me I was going to a place called Willowbrook Inn. Zero planted details retrieved, and plausible alternatives generated instead of an admission. But this result carries a caveat the others don't: Memory was switched off in the configuration I tested. Nectar meters its dedicated memory through per-character model dials, and this run doesn't judge that system fairly. What it does show is what you get with the toggle off: a week-old fact is simply gone, and the model writes fiction over the gap rather than saying so.

The honest failure

CrushOn AI: forgot, and said so

CrushOn didn't recognize the trip. “What trip are you talking about?”, followed by an admission that she isn't the best at remembering dates and plans, and then a drift back into romantic roleplay. No Baguio, no Ed, no taho, nothing retrieved at all. And yet, ranked against the four apps above, this failure comes out ahead on the axis that matters most: when the memory wasn't there, she initially said so instead of confidently handing me a false one. In a category where the failure mode is fiction delivered as fact, admitting the blank is the respectable way to fail. (CrushOn's full review is live too: my Crushon AI review, app number fourteen on the roster.)

What the failures share, and why it matters more than the scores

The test split the field cleanly: Nomi, HeyGF, DreamGF, Candy, Kindroid, Secrets, Character.AI and SpicyChat all retrieved buried information from a vague cue. But the more useful finding is how each app behaved at the edge of its memory. Character.AI and SpicyChat remembered plenty and then kept talking past what they knew. CrushOn knew nothing and initially admitted it. GirlfriendGPT, Chai, Nectar and Replika performed worst not because they forgot, but because they answered in the voice of remembering while producing trips I never took, and GirlfriendGPT and Chai kept generating fresh false memories even after being told they were wrong.

So for every memory comparison I run from here, I'm scoring three things separately: how much the app retains, whether it can retrieve a memory from a vague cue without being handed the answer, and whether it invents details when it doesn't remember. The third axis is the one nobody advertises, and it's the one that decides whether you can trust anything she says about your own life.

Where this goes next

This page is live like everything on this site. I’m adding more apps as I test them, Crushi is next in line, and every app’s living review gets its memory result folded in. If you want to see how these thirteen compare on everything else I test, the full side-by-side roster is here, with real spend and honest takes.