Applied research — consumer-scale AI companions
AI charactersthat remember you.
We build the memory, image and video systems that sit underneath consumer AI companions — the parts that make a character worth returning to.
- ~170
- Characters in catalogue
- Not yet published
- Images generated
- Not yet published
- Conversations held
- Not yet published
- Memories consolidated
Research that ships. Built to scale.
Most AI characters forget you the moment the context window rolls over. Ours do not. Lerk runs a dedicated memory subsystem that writes durable records from conversation, consolidates them into stable summaries over time, and retrieves them scoped to the character you are actually talking to.
It is a storage and retrieval problem, not a prompt trick. Memory survives the session, the device and the model.
- Persistent recall across sessions and devices
- Consolidation into durable, compact summaries
- Retrieval scoped per character, never leaked across them
- Correction and forgetting as first-class operations
A companion that looks like a different person in every image is not a character. Lerk runs a production image pipeline on Seedream via Novita, built around identity preservation — the same face, the same build, the same read, across poses, lighting and scenes.
The pipeline is designed for consumer volume: queued, moderated, and metered end to end.
- Identity held stable across poses and scenes
- Production pipeline on Seedream via Novita
- Moderation on the prompt and on the result
- Queueing and metering built for consumer volume
The same identity work extends into motion. Lerk generates short-form character video on Wan via Novita, sharing the character definition and the moderation path with the image pipeline so a character moves the way it looks.
- Short-form character video on Wan via Novita
- Shared character definition with the image pipeline
- Same moderation path as every other surface
- Asynchronous jobs with progress reported to the user
Underneath everything sits the part nobody sees: a routing layer that picks a model per turn, per-character system prompts assembled from the character definition and its memory, a metered credit economy, and moderation on both the way in and the way out.
It is the difference between a demo and something you can put in front of the public.
- Model routing decided per turn, not per app
- Per-character system prompts assembled from memory
- Metered credit economy across chat, image and video
- Moderation on input and on output
From research to product
Subsystems
Memory, identity and safety are built as independent systems with their own contracts — not as prompt engineering bolted onto a chat loop. Each one can be tested, replaced and reasoned about on its own.
RESEARCHInfrastructure
Routing, queueing, metering and moderation turn those subsystems into something that survives real traffic. Every generation is accounted for, every request is moderated, every model choice is deliberate.
SYSTEMSSurface
A catalogue of authored characters, tools for users to create their own, and an operations console for the people running it. The research only counts once somebody can use it without reading a paper.
PRODUCTBuilt to be fast. Designed to be cheap.
Companion products live or die on unit economics. A conversation is not one request, it is thousands over months, and every one of them carries an inference bill. Lerk treats cost as an architectural constraint rather than a line item to apologise for later.
- Model routed per turn — the smallest model that clears the bar
- Memory retrieved selectively, so prompts stay short as history grows
- Consolidation compresses history instead of replaying it
- Generation queued and deduplicated before it reaches a GPU
- Credits metered at the point of spend, visible to the user
- Moderation runs ahead of inference, not after the bill
The stack behind the product
- Inference
- Routed across frontier and open models, chosen per turn.
- Media
- Seedream for image and Wan for video, both served via Novita.
- Memory
- A dedicated store with consolidation and per-character retrieval.
- Platform
- Credits, moderation, catalogue and an operations console.
Why Lerk Studios in four answers.
Memory is the product
Everyone can call a chat API. Almost nobody has built the part that remembers. Long-term memory is the system we invested in first, and it is what keeps a Lerk character worth talking to long after the novelty has worn off.
One character, everywhere
The same character definition drives the conversation, the images and the video. Identity is preserved across all three rather than re-invented by whichever model happens to be running.
Built for consumer volume
Queueing, metering, moderation and routing were in the architecture before the first user. Consumer scale is a set of engineering decisions made early, not a migration performed under load.
Safety in the path, not beside it
Moderation sits on the input and the output of every surface — chat, image and video. It runs ahead of inference, which makes it cheaper to enforce and harder to route around.
Let's build together.
Partnerships, infrastructure, press, or a role on the team — tell us which and we will come back to you.
Or write to us directly
lerkbusisness@gmail.com