Founder and principal investigator at ZenSation AI, an independent research lab in Kiel, Germany, working on cognitive architectures for AI, agent memory and AI safety. Self-funded, without external grants or investors — income from pilots and consulting goes back into the research.
I have been putting software into production since 2009, across three domains: financial trading systems, enterprise CRM, and now AI research infrastructure.
zenbrain — a seven-layer, neuroscience-derived memory architecture for LLM agents. TypeScript with no third-party dependencies in its core packages, Apache-2.0. On LongMemEval-500 three of nine head-to-head comparisons hold against Letta, Mem0 and A-Mem — all three against A-Mem; the remaining six, against Letta and Mem0, show no significant difference (three competitors × three LLM judges, Bonferroni-corrected, version-matched) — and the measured arm (per the correction notice a Python port of a test-harness class, not the seven-layer system) reaches 91.3% of the accuracy of a full-context comparison arm (whose answer model reads the whole history) with 1/109.7 of its input tokens in the answer step.
voteworth — measures how often two voters of an LLM judge panel or voter ensemble pick the same answer when both break from the plurality, reports Kish's effective number of votes by raw agreement, and corrects the vote for voters that do, naming the panel shape where it cannot. Python, numpy only, Apache-2.0.
zenai — the self-hosted AI platform ZenBrain was extracted from, published as a frozen, readable snapshot.
| Preprint | arXiv:2604.23878 |
| Author ID | ORCID 0009-0001-1793-012X |
| Open records | Zenodo — CC BY 4.0, each with its own version and concept DOI |
| Reproduction | Mechanism ablation package — Apache-2.0; reproduces Tables 7–9 from a clean clone |
| Lab | zensation.ai · research@zensation.ai |
Every number I publish is meant to be re-derived, not believed. Benchmarks ship with the protocol, the effect sizes, and the cases where the system loses. Replications and counter-results are welcome — open an issue or write.


