Private, warm, and yours. Reach it like a friend, or open the app and talk, live, in real time.
The AI-companion market is real and funded. Replika, Character.ai and Nomi proved that millions of people want something to talk to that knows them. But the category split into two dead ends: romance-and-roleplay toys people outgrow, and clinical workbooks that never feel like a companion. Marrow is the third thing: a warm, private companion for adults that genuinely remembers you, and meets you both ways, as easy to reach as a text and as real as a voice on the phone.
Two things are true at once: general AI has no real memory of your life, and the apps that promised memory turned into toys people are now embarrassed to have on their phone. There's a mature audience with nowhere grounded to go.
ChatGPT, Gemini and Claude are task engines. Bolted-on "memory" is shallow and their context refreshes; Character.ai famously forgets the premise within a conversation. Nothing accumulates into knowing you.
Replika took a €5M GDPR fine and years of reputational damage. The "AI girlfriend/boyfriend" wave (Kindroid, Dusk) leans into roleplay and parasocial romance. It's exactly what mature users outgrow.
Woebot and Wysa are structured CBT workbooks, not companions. Anxiety-first tools like Stella are hyper-clinical. Useful, but they don't feel like someone who has time for you.
This isn't a product for lonely people. It's for people with full lives. Every human confidant comes with stakes: a relationship to you, a side to take, a verdict they'll keep. There's a whole class of things you carry that you can't hand to anyone you know. Marrow is the room with no stake, no side, and no verdict, that still remembers your whole story.
The person you'd normally process it with is the subject, or married to it.
"You can love him completely and still need to say something he can never unhear."Venting to a person leaves residue. Your bad day becomes their permanent, unconscious bias. You can't un-tell it.
"Say it here, so it doesn't land on someone who'll carry it."You pick your confidant by the verdict you want. Talking to friends is often asking a rigged jury.
"You don't need agreement. You need to hear yourself think."This de-stigmatizes the whole category and makes privacy intrinsic: the message literally is "it goes nowhere." It also arms the isolation defense: Marrow absorbs the vent so the relationship doesn't have to. It protects real connection instead of replacing it.
A warm, calm companion that remembers across every conversation and understands how the people and moments in your life connect. Reach it by text, by voice note, or by a live call.
Switch styles any time. The memory underneath stays the same person.
Warm and emotionally alive. The companion that genuinely knows you.
Reflective; holds space without judgement.
Direct and energising; helps you think, then move.
Ultra-brief. Just holds it.
Curious and deep; engages the idea itself.
As easy to reach as a text. As real as a voice on the phone.
Most companions ask you to live inside one app. Marrow meets you on your terms: start by texting it in a chat you already use, or open a native app built to be genuinely beautiful and simply talk to it. The same memory sits underneath, whichever way you reach in.
On WhatsApp, a deliberate call. WhatsApp is the obvious growth lever, but routing intimate conversations through it hands Meta the very thoughts people came to Marrow to protect. So the low-friction text channel is Telegram-first by design. A WhatsApp surface is on the table only with an architecture that keeps our privacy promise intact. We won't grow by quietly breaking the one thing this category can't afford to break.
Most "memory" is shallow vector recall: fetch a similar past sentence. Marrow also builds a private, per-user graph world-model, mapping the people, feelings and triggers in your life and the relationships between them. That structure is what lets it connect what you say today to what you said months ago, and grasp who and what is really at play.
Entities and KNOWS / FEELS / TRIGGERS edges extracted from every conversation. A living map of your world, not a transcript.
Emotions carry intensity and fade over time the way real ones do, so Marrow tracks how you feel now, not how you felt once, forever.
Every conversation deepens the model. The longer you stay, the more it knows and the harder it is to leave. Memory is the switching cost.
An app that "remembers everything" triggers an instinctive fear. So privacy isn't a policy page here. It sits next to every memory claim, and it is load-bearing.
Nomi and Kindroid store vast behavioural data to work. Replika was fined. That fear is the category's open wound, and our entry point with the exact audience that fled it.
Two axes decide this market: what the thing is for (roleplay and romance, or grounded reflection), and whether it truly remembers you (shallow recall, or a deep, structured memory of your life). Everyone clusters low and left. The corner that is both grounded and genuinely remembers sits empty.
Grounded reflection, warm but never flirtatious, with a deep, structured memory of your life underneath. No incumbent occupies it.
Frontier models are finally warm, multilingual and low-latency enough to feel like a person, and cheap enough per message that a monthly subscription still carries real margin.
The GDPR fines and the "AI girlfriend" backlash created a mature cohort actively looking for a grounded, private alternative. The demand is proven; the trust is up for grabs.
Real-time voice is now low-latency and natural enough to feel like a call with a person, not a robot. That turns live voice from a novelty into the thing people stay for.
The Live tier: full long-term memory, 5 companion styles, and real-time voice calls in 7 languages. Chat-only is CHF 5.99. A 14-day free trial runs on all three surfaces. Testing CHF 15 is the first pricing experiment this round funds.
The highest-ROI line in the plan, and it is built: 23% off monthly, collecting CHF 84 net on day one instead of over 18 months. Chat-only is CHF 59.99. Live on Telegram now; the iOS and Android builds carrying it are in store review. Annual plus card-on-trial is what takes LTV/CAC from 0.28x toward an expected 2.3x.
The entire production stack (Postgres, pgvector, Redis, workers, auto-TLS) runs on a single dedicated server. Burn is low by design; the model API is the main variable cost.
The switching cost is the accumulated memory. For a subscription, a compounding reason to stay is the whole game, and here it is the core of the product, not a bolt-on.
iOS App Store, Google Play, and Telegram, in 7 languages, from day one. The distribution surface is already built and shipping.
We're just out of beta: a production-grade product live across three surfaces and two app stores, with a real memory engine and privacy posture behind it. The platform is complete.
The single riskiest thing in this business isn't whether we can build it. It is built, encrypted, and shipping across three surfaces. The open question is how cold users convert at trial end. That is a unit-economics question, and we are answering it with high discipline.
Stated plainly, because diligence will find it anyway: today's trial still takes no card. That is the one lever we have not pulled, and we know exactly what it costs us. Two of the three are already done: we moved the trial from 3 days to 14 — our differentiator is memory that compounds, and that is legible at week three, not day three — and the annual plan is built and live, collecting CHF 84 net on day one rather than over 18 months. Card-on-trial is next, and it is a fortnight of work, not a rebuild. We know precisely which levers move our economics and we are pulling them in order.
We are already in motion: we are running a targeted, self-funded CHF 15,000 validation test to establish a cold-user CAC and a true D7/D30 retention baseline. The ask funds immediate product-economics optimization and scales only the channels that win this test.
Capital-efficient by construction: the product and infrastructure are done. Nearly every franc goes to proving and scaling the loop, not rebuilding the machine.
The product and infrastructure are built. This round funds product optimization (pricing & trial-ends), Swiss safety & crisis protocols, and a highly targeted scale campaign.
Apple Search Ads and Google Ads against what people actually type ("someone to talk to," "somewhere to vent"), landing on the store listing. Concentrated language bets to start, in German and Spanish, with every install attributable by promo code and UTM, and clinical or romance intent negated out to protect the brand.
Instagram and TikTok built on the "neutral room" idea: a truth about social life before it is an app pitch, so it earns organic reach. Warm creators whose audiences are strangers to us, plus an ambassador program, each carrying their own code so we see exactly who converts.
The onboarding economics that move LTV/CAC from a speculative 0.28x toward an expected 2.3x. Two of three are done and shipping — the 14-day trial and the annual plan. Card-on-trial completes it, before a franc of the test budget is spent.
A measured CHF 30k cold test yielding our first paying cohorts of strangers, establishing an honest CAC and a true D7 and D30 retention baseline.
CHF 180k concentrated solely in the exact channel or creator proven to convert cold users, turning Marrow from a product into a repeatable scale engine.
The place to say the things you carry around but never say out loud. It keeps your whole story, and answers to no one else.