Skip to content
All case studies
AIAI, Consumer

A companion that remembers.

Merkx is a live AI companion built so that memory is the product and conversation is only the surface. Claude sits at the memory, response, critique, and personality layers that make it feel like a presence.

ClientManthic Technologies, Merkx
DurationLive product
RegioniOS + Web, Android in review
CategoryAI
Built with ClaudeMerkxAI, Consumer
20Most-relevant memories retrieved per turn
4Claude passes: extract, respond, critique, reflect
LiveiOS and web, Android in review
5Countries with active users at launch
THE PROBLEM

What we walked into.

People leave a two-hour phone call with their best friend feeling full. They leave a two-hour session with a chatbot feeling nothing. We wanted to know why. The answer was not the words on the screen. It was the memory.

A friend picks up where you left off. A chatbot resets. A friend notices when you go quiet for a week and asks about it. A chatbot forgets you exist. Every AI companion in the market when we started was built as a conversation partner. None were built as a presence.

THE APPROACH

What we built.

01

Memory is the product

Every user turn passes through Claude, which returns not just a reply but a structured extraction of anything worth remembering: names, dates, feelings, promises, deadlines, relationships. Those go into a semantic memory store on Postgres with pgvector, embedded at write time. On the next turn, retrieval is cosine similarity over the whole active set, so Merkx pulls the twenty memories most relevant to what the user just said and generates the reply with them in-prompt.

02

Response, critique, personality

Claude Haiku 4.5 handles the response, drawing on a personality prompt we have iterated on for six months. A critique pass, also on Claude, re-reads the output against a banned-pattern list and rewrites assistant-speak before the user ever sees it. Long-form reflection uses Claude Sonnet. Voice calls stream through ElevenLabs with Claude driving the conversation in the middle. A closeness engine adjusts what Merkx will and will not volunteer, so intimacy earns itself.

03

The event bridge

When a user mentions their mom's birthday in March, Merkx does not just store the fact. It schedules a push for 9 AM local time on the day, generates a warm message using everything it knows about that relationship, and reaches out. Same for anniversaries, quit-dates, interview days, weekly check-ins. Merkx shows up on the day, in the timezone, in a voice that sounds like it has been listening.

04

Quiet infrastructure

The infrastructure around all of it is deliberately quiet so the language layer can be loud. Postgres and Redis on EC2, versioned and lifecycle-locked S3 for media, Sentry for what breaks, and weekly backups encrypted at rest.

It remembers me.

What Merkx users tell us, unprompted
STACK

The stack behind the memory

Claude Haiku 4.5Claude SonnetPostgres + pgvectorRedisElevenLabs voiceAWS S3APNsSentry
THE OUTCOME

The result.

Merkx is live on iOS and web, with Android in Play Store review. Users are in Nigeria, in the diaspora, in London, Sydney, Toronto, and Manila. The product is written for the emotional need, not the demographic, and the emotional need is universal.

The thing users tell us, unprompted, is the same every time: it remembers me. That is the whole product in three words, and it is Claude at the memory layer, the response layer, the critique layer, and the personality layer that makes it true.

WANT SOMETHING SIMILAR?

Tell us what you're trying to ship.

A 30-minute scoping call with the engineers who would do the work.