Safehouse Lab
§ 01 — The Lab

Agents that remember

Research and development on AI agents with persistent identity. Our own small models and neural networks, working with the large ones rather than hiding behind them — and the architecture that holds it all together.

§ The room, and the road

It packs into
two cases.

The lab packed for departure

Stephan works fully remote, between countries. The studio is not a desk you have to be sitting at — it is a tower, a panel and the engines on them, and all of it travels. Local inference does not care which timezone it is in.

That is also why so much of what the lab builds runs on its own hardware: a model you own boards the plane with you.

Packing the lab
At the lab
§ The base

Berlin,
Kreuzberg.

Safehouse Lab is based at betahaus in Kreuzberg — brick, glass and a lot of coffee, which is most of the reason this site looks the way it does.

It is the in-house lab for Vonfal AI and House of Ascania Industries: the research and the engines are built here, and the products ship under those names.

§ Main outposts
01 Berlin Germany
02 New York United States
03 Tel Aviv Israel
04 Puerto de la Cruz Tenerife, ES
01
Memory

Continuity is the product

An agent without memory is a search box with manners. We research persistent identity — what an agent keeps, how it distils a day into something worth remembering, and how that survives the software changing underneath it.

Memory is held as data the user owns, separate from the code that reads it, with additive-only migrations.

Persistent identity Distillation Data protection
02
Models

Small models, our own

We train and run our own small language models and neural networks rather than renting every thought. They work alongside the large models — the small one knows the house, the large one knows the world.

Running our own also means a cost floor we control and a privacy position we can actually stand behind.

Custom SLMs Local inference Fine-tuning
03
Architecture

Where the agent lives

Server and cloud architecture built for agents rather than adapted to them — the pod that holds a companion, the queue that feeds it, the storage that outlives any one model.

Zero-dependency where it can be, rented only where something genuinely bills us.

Server design Cloud Agent pods
04
Media engines

EMBER

Our in-house image, video and language engines, running on our own hardware. Campaign work, video and stills produced here rather than licensed per render.

Image Video Language
05
The commission

Ducal Research
Laboratory

houseofascania.com

Safehouse Lab serves as the Ducal Research Laboratory of the House of Ascania, first named in a charter of 26 October 1036. The House holds the charge; the lab does the work.

Two things sit inside that charge — the engine the record is kept in, and the historical research behind the Ascanian Chronicles: archives, editions and citations.

We do not verify the record. A source can be proposed by anyone who can read one, and nothing enters the Chronicles on this lab’s say-so. The House verifies. We build the instrument it is kept in.

CARTULARY Archives & editions Charter of 1036
CARTULARY A cartulary is the book a house copied its charters into. This one is the engine that holds the record — sources, editions, and the citation standing behind each line.
ASCANIAN CHRONICLES The research programme behind the record — and it names its own gaps on the front of it rather than writing around them.

On what we publish. We will tell you exactly what a product does, and we name any service your own data travels to. We do not publish which model sits behind a given agent, or how the studio is tooled — that is method, not a claim, and the architecture is the thing we actually sell.

Talk to the lab