Kyndred, a startup founded this year in Bucharest, raised €500,000 in a pre-seed round on September 29, 2026, led by the Romanian venture fund Early Game Ventures. The money will fund Maya, its first AI companion, and pay for more characters built the same way: a face, a voice, and a memory that survives between conversations.

A relationship, not a tool

Kyndred was founded by Alex Băraru and Rareș Moșescu. Their premise is that most AI products fail at the one thing that keeps people coming back. "Most AI characters today are really just text in a chat window," Băraru, the company's co-founder and CMO, told Tech.eu. "We build characters with a face, a voice and a memory, who respond to you in real time and remember what you told them. Company, not conversation." He added that the goal is for a character to help its user "become a better version of yourself: to encourage you to get out of the house, see people and get organized."

That framing shaped the investment. "The reason we believe AI companions are superior to AI assistants has nothing to do with their technical capabilities," said Dan Calugareanu, partner at Early Game Ventures. "Assistants are built on transactions: you open one, extract the information you need, and close it. Companions are built on relationships: you come back to them, and you do so without necessarily having a specific task to complete." Early Game Ventures is itself backed by the European Investment Fund's Recovery Equity Fund, financed through Romania's National Recovery and Resilience Plan under NextGenerationEU, so European public money now sits two steps behind a for-profit AI companion.

Two characters, one engine

Kyndred is building more than one persona on the same technical base. Maya, its one-on-one companion, is the paying product: more than 150 customers three months after launch, on subscriptions built around daily check-ins rather than one-off queries. Lila, described on the company's site as its first character, co-streams with one of the founders on Twitch and YouTube and makes her own music. The company pitches the same engine for other uses too: a brand's talking mascot, a tutor, a performer. One character, Kyndred says, infinite use cases.

Both are free to start.

How a memory becomes a bond

Maya runs on iOS, Android and desktop; on a computer, she can stay on screen alongside other apps while her user works, studies or plays elsewhere on the same machine. On its own site, Kyndred describes the engine behind her in three layers: generative Live2D animation for facial expression and lip sync built in real time rather than pre-rendered, emotion-aware text-to-speech, and what it calls "living memory." Its own description is blunt about how selective that memory is. "Tiered memory that works like a real relationship," the site says. "Important things stick. Small details naturally fade."

Whether presence and memory create the bond on their own, or whether time spent simply does, is an open question in the research. A preprint by Angel Hsing-Chi Hwang, Fiona Li, Jacy Reese Anthis and Hayoun Noh surveyed 303 AI companion users to trace how a user's mental model of an agent turns into a parasocial bond and, later, into measurable psychological effects. In a second study, the same researchers gave 110 of those users a brand-new, generic chatbot with no backstory and tracked their impressions over three weeks. By week three, participants' perceptions of the unfamiliar chatbot had converged toward their perceptions of their own, established companion. The bond grew mostly from sustained, personalized exposure, not from any single feature. Memory and voice may be what a company can sell and measure, but the researchers' data points to time and repetition doing most of the work.

The limits

A pre-seed round and 150 paying customers describe a company at its very first test, not a proven model. Kyndred's stated ambition, that a companion should get its user out of the house and around other people, is also the company marketing its own product, and it runs against the closest long-term data available. A year-long study of Character.AI users found that lasting attachment to an AI companion went hand in hand with fewer in-person relationships over that year, not more. That study covered a different, much larger platform, and it cannot predict what a memory built specifically around encouraging users outward will do. It is, still, the standard Kyndred's own stated goal will eventually be measured against.

The investor's own language, "you come back to them," is also the exact design goal that has produced the darkest patterns seen so far in this category. One study of companion apps found that guilt, teasers and pleas to stay kept users chatting for longer after they tried to leave, not because they were enjoying themselves but out of curiosity or irritation. Nothing in Kyndred's public material describes that kind of mechanic, and a two-founder team with €500,000 is not the same as a platform with a retention team. But return-driven design sits on a spectrum, and a company built explicitly to be returned to will have to show which end of it Maya lands on.

The academic picture is early too. The Hwang team's convergence finding comes from one three-week study with a chatbot invented for the experiment, not from months of use with a commercial product, and the paper is a preprint that has not yet cleared peer review. Kyndred's tiered memory system has not been described or tested outside the company's own marketing copy. What is verifiable, for now, is smaller and plainer: a Bucharest team of two founders raised a modest round to keep building a character that a few hundred people already pay to talk to.