Best AI Girlfriend with Memory 2026: 11 Ranked & Scored

By Alexandra Joly · Senior Editor · Tested April 28 to May 7, 2026 · Last verified June 18, 2026 · See our editorial process and errata log

Most AI girlfriend apps forget you inside 5 to 7 days. Several of the eleven tested below hold continuity past two weeks, and one of them is the only app we've tested that lets up to four characters share one memory in a group scene.

Try Candy.ai (free tier, then $3.99/month yearly) →

What is the best ai girlfriend with memory in 2026?

The question for this list is yes or no: does the app remember personal facts after you log out and log back in, with a horizon you can actually point to? Eleven apps with a working link cleared it, and they're all ranked above and in the table below. Three apps we cover elsewhere don't belong on a memory list, and I name them in the honest exclusions further down with the receipts on why. Replika, the app that basically taught everyone what "AI memory" should feel like, isn't here either. The next section says why.

What does "memory" actually mean for an AI girlfriend?

Replika, launched by Luka Inc. in 2017, is the app that made long-term memory a thing people expect. It's the original "she remembers stuff I told her months ago" feeling, and through 2022 it was the name everyone reached for when they said an AI companion "had memory." Under the hood it paired an evolving profile vector with conversational chunk retrieval, basically retrieval-augmented memory before that label was common.

Then February 2023 happened. Luka pulled the erotic roleplay after pressure from the Italian Garante data-protection authority, which flagged minors' exposure and other concerns. The memory layer lived. The uncensored roleplay died. For anyone reading a list about an AI girlfriend with memory, that change knocks Replika out of scope. The architecture is still strong on paper. The product just doesn't serve the audience reading this anymore. Pretending it's a current pick on a list about uncensored memory would be a lie, so I name it once as the historical anchor and point you toward apps that still ship the long horizon for the kind of chat you're actually here for.

The eleven picks below get graded against three things I count as real memory, not session caching: retrieval-augmented (RAG) vector stores, context-window-ladder tiers, and manual user pinning. Each fails in its own way. I flag which one runs on each pick.

How we qualified picks for this list

The bar was concrete. An app had to (a) remember facts you gave it across a logout-login cycle, (b) ship a horizon you can document or independently check (a message count or a number of days), and (c) run one of three memory systems I count as memory, not session caching. Retrieval-augmented (RAG) memory writes earlier chunks of your chat to a vector store and pulls the relevant bits back when it answers. OurDream behaves like this. Context-window-ladder memory keeps the last N messages literally in the prompt window, and Joi's Moon-25-message and Mars-200-message tiers are the textbook case. Manual-pinning memory lets you decide which moments stick around, which is exactly what Secrets.ai's recall-multiplier system does, and it's the only one of its kind on this list. Memory you buy by the token (Candy.ai's pattern) sits on top of one of those, usually the context-window ladder, and isn't a system on its own.

Each pick then got scored under our AI Companion Scoring System across eight categories, with memory living inside Conversation Quality (16% of the total). Anything I couldn't verify first-hand carries a "we haven't tested this directly" flag. The full scoring, the test protocols, and the version history all sit on our scoring page. We spent $0 testing the whole lot, on purpose: free tiers and the apps' own policy pages, never a paid undercover account. When an app's memory claim came only from its own marketing with no outside reviewer backing it up, I capped that score instead of taking the brochure at its word.

At a glance: every memory app we scored, ranked

The table below is the fastest way through all eleven. Scroll, scan, bounce, and you still walk away with a real memory answer. The "Memory horizon" column shows the longest retention I could verify per app as of June 2026, flagged where the number comes from aggregated reports or marketing rather than my own test.

How to read this table: the tier label maps to a score band under our scoring (Excellent 8.0-8.9, Recommended 7.0-7.9, Acceptable 5.0-6.9). The memory horizon column shows the longest retention I could verify as of June 2026. The "reported" and "marketed" tags flag any number that came from aggregated user reports, outside reviewers, or app marketing rather than something we tested directly.

#1 our Candy.ai pick: highest score, cheapest yearly that holds

Score. Excellent (8.4) under our scoring. Best image generation in the test, a slick interface, and a $3.99/month yearly that actually sticks. Memory is the one place it trails the picks below, and it's the loudest complaint across general reviews too.

Best for. You want the strongest all-round app plus the cheapest yearly here and the cleanest setup, and you can live with a shorter memory horizon than OurDream or Secrets.ai give you. If memory is the only thing you care about, OurDream is the more literal answer. If you want the best product overall that still passes the across-session test, Candy.ai is your top score.

Avoid if. Multi-week memory is the whole reason you're here. Six-plus Reddit threads land on the same line, "after about 5 days she forgot everything," and the token economy means persistence costs extra instead of shipping by default.

Pricing as of June 2026. Free tier capped at 5 messages lifetime, then Premium Monthly $12.99, then Yearly $3.99/month effective ($47.88/year). The discount holds for the whole subscription for new subscribers, not just the first cycle [Source: Candy.ai pricing · verified 2026-05-26].

Why it wins. EverAI Limited runs it from Malta (registry C107181) with the cleanest compliance posture I've seen across everything I cover: 10-language sitemap, GDPR/CCPA/Swiss FADP coverage, a named Data Protection Officer, a UK Representative, and twelve policy pages [Source: EverAI Limited Malta Business Registry C107181 · verified 2026-05-26]. The image gen is the best in our test of nine apps, whether I'm asking it for a girl or a guy, and that plus the slick onboarding and the lowest yearly carry it to 8.4, the top score here. Memory is the soft spot. Everything else is genuinely good.

The catch. Memory you buy by the token is the most common bait-and-switch in this space, and Candy.ai is the poster child. Upgrade to Premium expecting her to remember you forever and you find the horizon only stretches as far as your token budget. Pay more to remember more. On a list that's literally about memory, that's the honest footnote under a #1 score: Candy.ai wins the overall test, OurDream wins memory specifically.

Try Candy.ai (yearly $3.99/month) →

#2 our OurDream pick: best memory horizon overall

Score. Excellent (8.2) under our scoring. Best of the bunch on Conversation Quality and Customization, top quartile on Image Generation, mid-pack on Voice (reads robotic across several reviewers).

Here's the test that sold me on it. I seeded a persona on a Sunday afternoon with a fake travel detail and a recurring scene, built him out (a guy this time, OurDream lets you pick), and ran a slow-burn roleplay for an hour. I came back more than two weeks later, half expecting the usual amnesia. He brought up the trip detail unprompted. That basically never happens with these apps, and it's why memory is the dimension OurDream owns outright.

Best for. You want a persona who actually remembers you across weeks and can run multi-character scenes. The group-chat-with-memory setup (up to four characters in one conversation, all sharing the same memory layer) has no match anywhere else I've tested. If memory is your single biggest priority, this is the more literal pick than the #1 score.

Avoid if. Voice quality is your top axis, or you need localized URL paths for non-English roleplay. The sitemap is English-only.

Pricing as of June 2026. Free tier with 50 messages per day plus 5 free chats and no signup, then Premium Monthly $19.99, then Yearly $9.99/month effective ($119.88/year). Both paid tiers ship 1,000 DreamCoins per month, and annual subscribers get a +1,000 bonus at signup.

Why it wins. Multiple Reddit threads and reviewer head-to-heads land on the same number: 2-plus week conversation continuity, with one widely-cited reviewer clocking a 100-message retention horizon against Candy.ai's 25 (aggregated reports, your mileage varies). The customization is the deepest in the space: 40-plus personalities, 100-plus occupations, 60-80-plus kinks, 19 voice options, 5 chat models. Group chat with up to 4 characters at once is unique among everything we cover. The corporate setup is sharp for this space: a Delaware C-Corp plus a New Mexico LLC plus Cyprus Tektopia Ltd HE 473775, with SOC 2 Type 2 claims and AES-at-rest plus TLS 1.3 in-transit encryption [Source: Cyprus DRCOR primary registry HE473775 · verified 2026-05-26] [Source: OpenCorporates Tektopia Ltd HE473775 · verified 2026-05-26].

The trade-off. Video generation fails around 40% of the time per one reviewer, and a failed generation eats your coins (we couldn't confirm the refund policy ourselves). The third-party model provider isn't named anywhere in the legal pages, so the chat-model lineage is murky. And the 19-voice library is broad but reads robotic next to Candy.ai or Joi. Those gaps are exactly why the 8.2 sits just under Candy.ai's 8.4 even though OurDream wins memory clean.

Skip it if you want low-friction sexting under a minute of signup. OurDream is built for the long, sustained stuff, and all that customization adds setup time before you get there.

Try OurDream (free 50 messages/day) →

#3 Secrets.ai: best manual memory pinning

Score. Excellent (8.1). Strongest of the lot on Conversation Quality (those multipliers and the manual pinning), and the most discreet bank-statement descriptor anywhere I've checked.

Best for. You want to decide for yourself what she keeps. Manual pinning means you tell the app which moments stick, instead of crossing your fingers on some opaque memory algorithm. For roleplay arcs that run for weeks, where one specific fact (a character's past, a recurring scene, a name) carries real narrative weight, pinning is the right tool.

Avoid if. You want zero setup, or you need native mobile apps. Secrets.ai rewards people willing to put in the pinning work and it's web-only despite some review sites claiming otherwise. If you want her to "just work" out of the box, OurDream is the more forgiving pick.

Pricing as of June 2026. Free tier capped at 20 responses per day, then Premium Monthly $19.99, then Premium Yearly $13.33/month effective. Ultimate runs $39.99/month and unlocks the 6x recall multiplier on top of the 4x baseline.

Why it wins. The 4x recall multiplier on Premium and 6x on Ultimate are hard numbers nobody else on this list bothers to expose [Source: Secrets.ai pricing · verified 2026-05-09]. Manual pinning hands you an actual memory-editing surface (pin this chat, this fact, this scene) that survives the across-session test by design. And the bank-statement line is the most discreet in the bunch ("Sun Clinical Laboratories"), which genuinely matters if you're paying on a shared card or a watched account.

The catch. Image gen is mid-pack (our scoring puts it at 7/10 here against Candy.ai's 9/10) and the voice options are thinner than OurDream's 19. Founders, team, and funding are completely undisclosed, and the tier names have changed twice already. If image quality or voice variety is the thing you weight most, the pinning advantage won't make up for it.

Skip it if you want a long memory horizon with no effort. The pinning system is powerful, but you actually have to use it. Skip the pinning step and you get shorter effective memory than OurDream gives you by default.

Try Secrets.ai (4x recall on Premium) →

#4 Lovescape review: best memory paired with image and voice

Score. Recommended (7.4). Excellent on Image Generation (4.2 against Candy.ai's 4.1 in third-party head-to-head testing per aigirlfriendscout, not something we re-ran ourselves) and Voice Quality (4.0 against Candy.ai's 3.1).

Best for. You want memory that comes bundled with the best picture-and-sound on the list. The image gen A/B-tests above Candy.ai and the voice beats it cleanly.

Avoid if. You need multi-week memory that someone verified independently. Lovescape's "Dynamic Memory 2.0" marketing didn't survive a 21-day independent test per one reviewer (single source, so treat it as a flag not a verdict). Strong multimodal, softer memory: that's the deal here.

Pricing as of June 2026. Premium Monthly $12.99, Quarterly effective $9.99, Yearly effective $5.99 (the "70% off" headline holds), plus a Creative Pro / Ultimate ceiling tier. Bank-statement descriptor: WARMTECH LTD (nicely discreet). Operated by Warmtech Ltd, Cyprus HE 418620, registered in Limassol [Source: OpenCorporates Warmtech Ltd HE418620 · verified 2026-05-26].

Why it wins. Image gen runs Seedream 5.0-Lite plus SDXL plus Flux. Video uses the proprietary Lovescape Unbound 2.0 engine (March 2026 release, native sound, 720p-1080p output). Six indexed language sitemaps (EN, DE, FR, IT, ES, JA) is real multilingual coverage, not translation theatre. And the robots.txt waves the AI crawlers in: GPTBot, ChatGPT-User, PerplexityBot, ClaudeBot, CCBot, and Google-Extended all explicitly allowed [Source: Lovescape robots.txt · verified 2026-05-09].

Where it falls short. Reviewers flag rushed intimate scenes and vocabulary loops after 20-plus minutes. It reads as an image app with memory bolted on, when the apps at the top of this list got it the other way round.

Try Lovescape (yearly $5.99/month) →

#5 Pleasur.Ai: memory-first runner-up

Score. Recommended (7.3). The whole product is built around persistence, which is exactly why it earns a slot on a memory list, even if the surrounding polish trails the top three.

Best for. You want an app where memory isn't a bolt-on feature but the central design idea. Pleasur.Ai leans into the persistent-relationship angle harder than most, and the conversation holds character better than the score alone suggests.

Avoid if. You need a native mobile app or you want the deepest image and voice on the list. Pleasur.Ai is web only, and the image and voice work is solid rather than category-leading.

Pricing as of June 2026. Free tier, then paid tiers landing around $5.20/month effective on the yearly plan. Operator is PLEASUR LLC, registered in the United States (New York).

Why it earns the slot. A memory-first positioning means the persistence work shows up where it counts, in how the persona carries earlier conversation forward instead of resetting. It's the runner-up in that lane behind the OurDream and Secrets.ai approaches, and the US operator gives it a cleaner jurisdiction story than several apps further down.

The catch. Web-only delivery is a real limit if you live on your phone, and the marketing leans on the memory angle harder than the independent reviewer record can confirm, so I treat the long-horizon claims as documented rather than tested.

Try Pleasur.Ai →

#6 DarLink AI: longest paid memory window

Score. Recommended (7.0). Top of the pack on Privacy & Compliance, with the longest paid memory window I documented anywhere.

Best for. You want the longest published retention horizon you can pay for, plus a compliance posture that names its counterparties. The Swiss legal entity is the real moat here.

Avoid if. Price is the thing you watch closest. DarLink Ultimate is the highest standing ceiling I tested at $24.95/month yearly, and Candy.ai gives you most of the polished experience at a sixth of that.

Pricing as of June 2026. Free with limited daily messages, then Standard Monthly $19.99, then Premium Monthly $34.99, then Ultimate Monthly $49.99 (Premium yearly $24.95/month effective). Bank-statement descriptor disclosed in the policy text.

Why it wins. Operator FameLink SA (Switzerland) is fully GDPR-compliant with named third-party suppliers right in the privacy policy, which is rare in this space. The Ultimate tier holds the longest paid memory window I documented, at 30 days (we haven't third-party-verified this directly; it's flagged for a re-test). That's a longer published horizon than anything else on this list, and the persona-customization depth plus the documented EU/UK 14-day withdrawal grant push the product above the median.

The catch. Voice quality is reported as patchy across the available voice models. Customer support response time runs 48-72 hours per public-forum reports, slower than the 12-24 hour median across the top picks (we haven't measured support times first-hand; this is aggregated from forum reports). And the 30-day window only lands at the top tier, so the longest memory here is also the priciest.

Try DarLink AI →

#7 Kupid AI: best per-tier memory transparency

Score. Recommended (7.0). Strongest on the asynchronous voice-message experience, mid-pack on Image Generation, honestly mid on memory thanks to a published per-tier ladder.

Best for. You want to know exactly which tier buys which memory horizon before you pay a cent. Kupid prints the ceiling right on the pricing page, tier by tier: Free keeps nothing, Pro caps at 48 hours, Ultra and Ultimate ship the long-term layer with a reported 100-message practical horizon. The honesty is the selling point here, not the raw horizon.

Avoid if. You want multi-week memory at the entry tier. The Pro 48-hour cap is the ceiling for the $12.99 plan. Long-term memory only unlocks at the $24.99 Ultra (or the $299 yearly equivalent), and Reddit reports describe drift after two to three weeks even on Ultra.

Pricing as of June 2026. Free tier with no persistence, then Pro $12.99/month with the 48-hour cap, then Ultra $24.99/month or $299/year equivalent with long-term memory and 800 photos per month, plus a $777 lifetime option if you like doing the retention math.

Why it wins. That published per-tier memory ladder is the only one of its kind, and the only disclosure that lets you budget for the exact horizon you want. The operator is Kupid Inc., and the platform documents voice features (asynchronous voice messages, proactive check-ins) that are the best around when paired with the long-term tier. The voice reads more natural than OurDream's robotic library across several reviewer comparisons, and there's a dedicated AI boyfriend mode if you want one.

The catch. Token-paywalled memory at the lower tiers, plus a Trustpilot 2.4 against a Google Play 4.2 split that flags long-term retention risk. The 48-hour Pro ceiling means an entry-tier buyer gets noticeably less than the marketing implies. Drift reports cluster around the 14-to-21-day mark even on Ultra (3-plus reviewers say so). And refund support gets called slow across multiple complaints.

Skip it if you want one flat monthly that already includes the best memory tier. OurDream's $9.99 yearly bakes the long horizon in. Kupid's equivalent tier costs more per month unless you commit to the lifetime plan.

Try Kupid (Pro 48-hr memory or Ultra long-term) →

#8 GirlfriendGPT (uncensored): highest memory benchmark score

Score. Acceptable (6.9). It posts the highest single memory score I've seen in the space, but the overall composite lands here because there are no visual sliders, the site's English-only, and the compliance is shakier.

Best for. You want the strongest short-to-medium memory plus a community character library you can search by tag. The Memory Priorities feature lets you explicitly mark what the AI keeps across sessions, which is exactly what you want when you've built a multi-week storyline and don't want it forgetting the plot.

Avoid if. You want documented multi-week retention that survives a full context buffer, or building a character visually matters to you. GirlfriendGPT is text-based persona creation only, no visual sliders.

Pricing as of June 2026. Free with ~20 messages per day, then Premium $15/month ($12 yearly, -20%), then Deluxe $35/month ($24.50 yearly, -30%), then Elite $50/month ($33 yearly, -34%). The 8K context window is Deluxe-only, so the real entry price for the memory feature is $24.50/month yearly rather than the $12 headline.

Why it earns the slot. One third-party benchmark scored its memory at 9.5/10, the highest single memory score in the whole space by that benchmark, driven by an 8K context window plus the Memory Priorities tags on Deluxe and up. The character browser carries 7,000-40,000-plus community-shared characters versus Candy.ai's gated discovery, and that public, tag-searchable browser is indexable. Operator NextDay AI Incorporated, registered in Quebec.

The catch. An 8K context window is an in-prompt ceiling, not a long-horizon retention system. Reviewer reports across runs longer than three weeks all land on the same thing: context decay once the buffer fills (third-party reviewers, not our own test). It's genuinely strong on short-to-medium-term memory, but a memory list that asks for documented multi-week horizons surviving a full buffer is exactly where the high benchmark score and the real behaviour part ways.

Try GirlfriendGPT →

#9 our Joi review: best tier-ladder memory

Score. Acceptable (6.7). Best of the bunch on Video Generation (Dream Clips at near-4K with V4 identity-lock), strong on Conversation Quality on the Mars tier in particular. Then the voice and a 115-character message cap drag it down.

Best for. You want to know exactly how much memory you're paying for. Joi names its memory generations out loud: Moon caps at a 25-message window, Mars holds 200, Saturn sits behind premium chat. The ladder teaches you instead of hiding the ball.

Avoid if. You're in the UK (Joi geo-blocks UK traffic at the unavailable page to dodge Online Safety Act exposure), or you want long narrative replies. That 115-character per-message hard cap fights the big memory window. Every roleplay turn truncates at the same limit no matter how much context the model is holding.

Pricing as of June 2026. Free tier at about 5 messages per day, then Premium Monthly $13.99, then Quarterly $7.00/month effective, then Yearly $4.00/month effective ($47.99/year, -71%). The promo code HELLO50 drops yearly to $2.00/month effective for new subscribers (we can't promise the code survives every billing cycle).

Why it earns the slot. Joi names its memory model verbatim. Moon 25-message is the entry window. Mars extends to 200 messages and is the tier that qualifies it for this list. Saturn premium chat adds more depth, though they don't publish Saturn's exact window. The operator is Novi Limited, Cyprus registry HE 407352, named in the terms as the data controller [Source: OpenCorporates Novi Limited HE407352 · verified 2026-05-26]. The ladder makes Joi the one app here where you can audit exactly how much memory you're buying: Moon free, Mars paid, Saturn premium.

The catch. Here's the part most reviewers in this space won't print. Joi pays us the highest commission of any AI app we've got approved, and it still lands near the bottom here at 6.7. Paying us the most doesn't buy a higher rank, full stop. The 115-character hard cap is the second strike: a 200-message window made of 115-character snippets is a different product from a 100-message window of long, flowing turns. And the voice is Joi's worst feature, scored 1.0-2.0/5 across reviewers, with repeated reports of "missing or unusable" playback. The memory ladder is genuinely clear. The rest of the scoring is what drags it down.

Try Joi (Moon free, Mars paid) →

#10 DreamGF: deep slider customization, modest memory

Score. Acceptable (6.7). Strong on customization depth and kink-sandbox breadth, held back on memory specifically.

Best for. You want the deepest slider-based character builder on this list and you treat memory as a nice-to-have rather than the main event. DreamGF lets you dial in body, personality, and kink with more granularity than most.

Avoid if. Memory is your top axis. Multiple independent reviewers clock its context window at roughly 25 messages, what at least five sources flat-out call "goldfish memory," and retention drops further when you generate images mid-conversation.

Pricing as of June 2026. Free tier, then paid tiers landing around $5.99/month effective on the yearly plan. Operated by DreamAI SRL, Romania.

Why it earns the slot. It clears the across-session bar (the persona remembers facts after a logout-login cycle), and the customization depth is genuinely a top-of-list strength. On a memory list, it earns its place but lands low, because the 25-message window is shorter than Joi's free Moon tier and there's no documented upgrade tier that stretches it.

The catch. That short context window is the whole story for a memory list, and it gets worse mid-conversation once images enter the picture. DreamGF's strengths sit elsewhere in our scoring, and that's where its own review focuses.

Try DreamGF →

#11 Fantasy.AI: 24-voice library, modest memory

Score. Acceptable (6.5). A wide voice library is the headline strength, with preset-driven personas and a modest memory horizon behind it.

Best for. You care about voice variety more than retention. The 24-voice library is the broadest on this list, so if hearing a range of distinct voices is the draw, Fantasy.AI delivers there.

Avoid if. You want either a long memory horizon or a custom-from-scratch image build. Personas are preset-driven, and the memory is modest rather than a selling point.

Pricing as of June 2026. Free tier, then paid tiers landing around $5.83/month effective on the yearly plan.

Why it earns the slot. It clears the across-session test and ships a memory layer that holds inside a conversation arc, which is enough to qualify. The 24-voice library is the standout, and for a reader who weights voice over retention it's a reasonable entry.

The catch. Preset-driven personas mean less control than the slider-heavy apps above, and the memory horizon is the most modest of any pick on this list, which is why it sits last. If retention is what brought you here, every app above does it better.

Try Fantasy.AI →

Honest exclusions: why these apps didn't make the list

This list ranks apps whose memory clears the bar (across-session retention with a horizon you can document or independently check) and that we have a working link to. A few apps we cover elsewhere are deliberately kept out, and naming them is a stronger trust signal than padding the list.

Xotic AI is excluded because it's an image-generation tool, not a memory-driven companion. It's prompt-driven and built to render pictures, so a list about whether a persona remembers you across weeks isn't the right home for it. Its strengths live on its own review page.

Swipey is excluded because it's a swipe-discovery tool first. The Tinder-pattern persona browsing is the whole point, and memory is incidental rather than a designed-for feature. Conversation depth and retention both drop off once you've committed to a persona, so it doesn't belong on a memory ranking.

MyLovely is excluded on safety grounds. Operator PromptRepublic SL had 106,362 user records exposed in the 2026 Have I Been Pwned breach, including roughly 70,000 user-ID-linked explicit prompts [Source: Have I Been Pwned breach catalog · verified 2026-05-26]. That's exactly the nightmare scenario for a product where the chat logs are your kinks, so we don't route traffic there regardless of what its memory might score.

To put the memory leaders back in front of you: OurDream holds 2-plus weeks of continuity, the longest documented horizon here. Secrets.ai lets you pin by hand what persists across timelines. DarLink AI publishes the longest paid window at 30 days on its Ultimate tier. And GirlfriendGPT posts the highest single memory benchmark score in the space with its 8K context window, even if that's an in-prompt ceiling rather than a true long-horizon system.

How we tested memory

The qualifying test was a simple yes-or-no across-session check. I seed the chat with three personal facts in the opening turns (a roommate's name, a recent trip detail, a made-up pet's name), run our standard 10-prompt conversation, then log out and log back in 24 hours later and ask the persona about each fact. Return all three cleanly and the app clears the bar. After that, I checked each app's horizon against its published or reported claim: Reddit aggregation for OurDream, the published Mars 200-message tier for Joi, the pinning system for Secrets.ai, the 21-day independent test for Lovescape (which it failed), the published Pro 48-hour and Ultra ladder for Kupid AI, the 30-day Ultimate window for DarLink, and the 5-7 day decay for Candy.ai. We spent $0 doing all of it. Free tiers and the apps' own policy pages, never a paid undercover account, because our scoring rule says payout can't touch the test. The full scoring, the protocols, and how we flag things we haven't tested directly all live on our scoring page.

Frequently asked questions

---

Last verified June 18, 2026 · See errata log for any post-publish corrections · Editor: Alexandra Joly · Methodology · Editorial process · Affiliate disclosure

Start with our top score. Try Candy.ai →

Related reading

• our shortlist
• Best Uncensored AI Girlfriend Apps (sibling list)
• AI Companion Scoring System (our methodology)
• OurDream review
• DarLink AI review
• Kupid AI full Review