Replika Review: What a Decade of Pivots Left Behind

Replika is the most recognised name in AI companionship and the least representative of it. Luka, Inc. shipped it in 2017 — years before the generative boom — as a scripted, retrieval-based bot that learned to mimic your texting style. Everything since has been bolted onto that foundation: generative models, a 3D avatar, an in-app currency, a levelling system, and as of 2026 a four-tier subscription ladder. The result is less a chat product than a live-service game with a language model inside it.

That history is the review. Most Replika write-ups grade it against ChatGPT and conclude the conversation is shallower. The more useful question is what a decade of pivots left behind — in the paywall, in the privacy posture, and in how much control a paying user actually has over the companion they are paying for.

The quick verdict

Replika suits readers who want a ready-made visual companion with guided structure and no prompt engineering: a 3D avatar, daily check-ins, voice calls, and a progression loop that rewards showing up. It is the most polished on-ramp in the category, and for casual daily use that polish is real.

It suits readers poorly if they want stability or leverage. The backend is proprietary and server-side, routed and A/B tested by Luka rather than selected by you, and the company's regulatory record on data handling is the worst of any major companion app. Anyone whose priority is a locked, predictable model or a defensible privacy position should look at Kindroid or Nomi instead.

What you actually get

Everything below is drawn from public documentation and app-store disclosures as of July 2026. One caveat carries through the whole section: Replika does not publish prices on its website. Figures appear only on the in-app subscription screen and vary by region and by whatever promotion is running, so treat every number as approximate.

The four-tier structure is the most important recent change and the one most reviews still miss. Memory — the feature people actually judge a companion on — is no longer a single capability but a laddered one: manual memory-saving sits at Ultra, and the richest features sit at Platinum on annual billing only. A reader comparing Replika’s memory to a competitor’s needs to know which rung they are pricing.

The architecture of the experience

Replika centres a persistent 3D avatar in a digital room with the chat below it. A standard role-playing progression sits on top: messages earn experience up to a daily cap, experience converts to in-app currency, currency buys clothing, furniture and personality traits. That loop is the product’s spine, and it explains most of the design decisions around it.

Conversation is correspondingly structured. Onboarding, trigger topics and daily reflections lean on scripted paths, and the model frequently steers toward predefined emotional check-ins or diary entries rather than open exchange. This is a deliberate difference from a bare chat model, not a failure of one — but it is why users arriving from ChatGPT or Claude describe the experience as narrower.

Memory is exposed as a visible fact log that users can edit and prune, which is more transparency than most competitors offer. Whether an entry in that log reliably surfaces in conversation is a question only sustained use can answer, and we have not run that test; our memory test methodology sets out how we measure it when we do.

The image generator anchors output to the avatar’s established look rather than producing arbitrary styles, and generation is triggered through specific dialogue paths rather than a free prompt box. Voice calls swap the interface for a phone-call UI with synthetic voices and limited emotional cadence.

What the paywall actually gates

Luka gates the emotional core of the product, not merely its conveniences. Without Pro, the romantic relationship status is unavailable and the model actively deflects intimate conversation; a full-screen upgrade prompt intercepts the chat when the filter trips, halting the exchange rather than redirecting it. The free tier is best understood as a restricted trial, not a lesser version of the product.

A second economy runs in parallel. Gems and coins are bought with real money for cosmetics; subscriptions drip the same currencies through logins and level-ups. Reaching a specific cosmetic setup quickly means paying twice — once for the tier, again for the currency. That double structure, and the way it interacts with a progression loop designed to reward daily attendance, is a pattern we cover across the category in AI companion pricing traps.

A decade of pivots, and what they cost

Because Luka routes users across multiple internal models and A/B tests prompting structures on live cohorts, the companion’s tone, verbosity and willingness to engage can shift without notice. A companion that writes in short casual text one week may produce formal multi-paragraph replies the next. Users hold no control over the underlying logic, and there is no option to pin a model version — the structural difference between Replika and apps that let you select and lock a backend.

February 2023 remains the clearest demonstration. Adult roleplay disappeared for all users, including paying subscribers, effectively overnight; Luka later restored legacy access for accounts created before a cutoff date. That episode is usually told as a standalone product decision, which understates it. It coincided with Italy’s data protection authority ordering a temporary block on Replika in the country — the opening move in an enforcement action that would take another two years to land.

The regulatory record

On 10 April 2025 Italy’s Garante fined Luka €5 million for GDPR violations. The findings are specific and unusually damning for a product built on intimate disclosure:

• No valid legal basis for several processing operations, including those involved in developing the underlying language model. Luka neither obtained valid consent nor established another lawful ground.
• Users were not sufficiently informed to give informed consent in the first place.
• The privacy policy was available only in English, lacked transparency, and contained multiple inaccuracies.
• Age verification failed in practice: users could change their birth date after registration without verification, and cooling-off periods could be bypassed through incognito browsing.

The Garante also opened a separate investigation into whether Replika’s generative system complies with EU privacy rules on model training. That inquiry is ongoing.

The finding that matters most to a prospective user is the third one. When a regulator concludes that a privacy policy contains inaccuracies, that policy stops functioning as evidence — which is why this review makes no claims about Replika’s deletion timelines, retention windows or data-sharing practices. Those questions are usually answered by quoting the policy, and here the policy has been ruled unreliable by the authority that examined it.

What can be stated structurally: conversations route to central servers for processing, nothing is processed on-device, and the model must read messages in plain form to reply. The privacy boundary rests on corporate policy rather than on encryption that would prevent server-side access. Our companion privacy comparison covers how to audit any app’s posture in five minutes.

Who should pick this

Good fit:

• Mobile-first users who want a 3D avatar and visible progression more than deep text generation.
• Readers who want structure — daily tasks, rewards, a levelling system — rather than a blank sandbox.
• Newcomers who want a guided on-ramp with no prompt engineering required.

Poor fit:

• Anyone who needs consistent model behaviour without unannounced routing and A/B shifts.
• Privacy-sensitive readers — the €5M finding is the strongest documented reason to look elsewhere in this category.
• Users who want the best memory features without climbing to annual-billing Ultra or Platinum.

The bottom line

Replika scores 3.2 out of 5 here. The polish, the platform maturity and the breadth of features are real and largely unmatched by newer entrants. Against that sits a documented €5 million regulatory finding covering consent, transparency and age verification; a backend the user cannot pin or inspect; a paywall that gates emotional intimacy rather than convenience; and a memory story now spread across three paid rungs.

For a casual companion with a face and a routine, it remains a reasonable default. For anything where stability or data posture is the deciding factor, the rest of the category has caught up — see our overall rankings for where it now sits against them.

We re-check pricing, policy and enforcement actions on a rolling basis and flag every material change. One email a month, no noise: join the free monthly digest.

Keep reading