Independent AI Research Lab

Our research — across three tracks

Independent fundamental and applied research: open research cores such as ZenBrain — and the application fields built on them: applied AI for knowledge work (Eigenraum by ZenSation) and safeguarding public spaces (CrowdGuard). Self-funded, open access, with civil-liberties architecture as a design principle.

Application fieldsTRACK BSafety in Public SpacesTRACK CApplied AI for Knowledge WorkFurther domainsZenBrainCORE · TRACK A
ZenBrainCore · Track A

Application fields

TRACK BSafety in Public Spaces
TRACK CApplied AI for Knowledge Work
+ Further domains

Core & application fields

One core, two application fields

FoundationTrack A

Cognitive Architectures

How AI remembers, reasons, and learns over time. Seven-layer memory architecture (ZenBrain), sleep consolidation, neuromodulator-driven consolidation, FSRS-style review. The shared core that Tracks B and C build on.

Research outputs

  • ZenBrain — open-source core on npm (Apache 2.0)
  • Open-access papers on arXiv and Zenodo (DOI)
  • 9 foundational algorithms and 6 PMA components
Track A in detail →
Track B

Safety in Public Spaces

Research on the early detection of critical situations in public spaces and at mass-casualty incidents — explicitly without biometric identification. Three-outlier model, helper protection as an objective, civil-liberties architecture in eight mandatory corrections.

Research outputs

  • CrowdGuard — research prototype
  • Civil-Liberties position paper in preparation
  • OpenTimestamps-anchored research documents
Track B in detail →
Track C

Applied AI for Knowledge Work

Eigenraum as the AI operating system for the enterprise: consolidating data streams, generating insights and alerts automatically, business intelligence — plus AI-augmented meetings, project management, and process mapping. How organisations adopt AI step by step, with humans at the centre.

Research outputs

  • Eigenraum — AI operating system for the enterprise, a productive research application (the extended version of our first product, ZenAI)
  • Process-atlas methodology (in research)
  • Three-year AI adoption framework (methodology in development)
Track C in detail →

Research principles

What we commit to

Open access

Research outputs are published on arXiv and Zenodo with DOI — freely accessible, permanently archived.

Forensic integrity

Core research documents are anchored in time via OpenTimestamps. This produces a cryptographically verifiable proof of time: what is proven is that a document existed in this form at that moment — independent of later publication dates.

Civil-liberties by design

Safety research with fundamental-rights grounding — explicitly without biometrics, with helper protection and the safeguard of political assemblies as structural design decisions.

Research with application

Algorithms are not only published — they run in our own operation (Eigenraum, since August 2026). Research has to hold up.

Methodological standards

How we work in practice

Pre-registration, reproducibility, external validation, documented ablation, and data minimisation are the operational standards against which we hold our work accountable from the first hypothesis to publication — documented in detail on the methodology page.

Operational detail on the methodology page

Origins

How this research grew

A three-year arc — from the first conceptual sketch to architectural consolidation to open-access publication. The milestones below are documented in the internal development repository and in the publications themselves.

  1. 2009 – 2023

    Groundwork. Algorithmic work on statistical models, deterministic decision systems, and real-time processing. In parallel, several years of digitalisation and transformation projects — a twofold preparation that later shows in the architecture.

  2. Mid 2024

    The idea for a cognitive architecture with real long-term memory takes its first form. First sketches of a multi-layer memory structure drawing on neuroscience (Ebbinghaus, FSRS, Hebbian learning).

  3. 2024 / Early 2025

    Architectural principles take concrete shape: seven memory layers, sleep consolidation as a methodological core, Bayesian confidence propagation. In parallel, the civil-liberties architecture for the public-safety track matures.

  4. Mid – Late 2025

    First production implementations of the memory layers. Build-out of the test suite, methodological pre-registration via OpenTimestamps. Beginning of the research documentation as an independent source.

  5. January 2026

    Start of the production implementation phase in the current internal development repository. Eight consecutive architecture phases in the first two weeks.

  6. March 2026

    Consolidation as an open-source project. ZenBrain is extracted as a standalone package family and released on npm under Apache 2.0.

  7. 31 March 2026

    The first version of the ZenBrain paper appears openly on Zenodo (DOI 10.5281/zenodo.19353663).

  8. April 2026

    Academic preparation: LaTeX paper infrastructure, ORCID profile, arXiv endorsement. On 26 April the preprint appears on arXiv (2604.23878), open access.

  9. Now

    Preparation of follow-up publications for peer-reviewed venues. Exploratory exchanges with universities and research consortia across the European research area.

Reception

Documented early, placed by others

ZenBrain describes a memory lifecycle for AI agents with explicit, active forgetting; its first version has been openly available on Zenodo since 31 March 2026.

The survey "Agentic Evolution" by the Sico team at Microsoft Research reviews about 300 papers and calls forgetting in agent memory "critically understudied": very few of the roughly one hundred papers in its Memory & Sense group implement it explicitly. ZenBrain is one of the four works in the sub-group "Forgetting/lifecycle" and the survey's only source for sleep-based consolidation.

In MindMemOS, Huawei's Noah's Ark Lab grounds the foundational sentence of its introduction in three sources: two surveys and a single system paper, ZenBrain.

We see ZenBrain as an early bridge contribution to dynamic, lifecycle-oriented memory architectures: independent, externally recognised and openly tested.

All citing works and their criticism

Research agenda

What is next

Exploratory work and preparatory steps — kept strictly conditional until written consent or binding funding decisions are in place.

Horizon: 2026 – 2027

Track A — Cognitive architectures

  • Follow-up publications focused on memory consolidation and sleep replay (peer-reviewed venues, in preparation).
  • Stable releases of the Postgres and SQLite adapters for ZenBrain (Apache 2.0, npm).
  • Replication packages with full eval scripts and datasets on Zenodo.

Track B — Safety in public spaces

  • Civil-liberties position paper on a fundamental-rights-preserving safety architecture (in preparation).
  • External validation of the three-outlier architecture via Notified-Body pre-assessment and an independent legal opinion.
  • Exploratory exchanges with research consortia on proposals along BMBF SIFO and Horizon Europe Cluster 3.

Track C — Applied AI

  • Methods paper on the process-atlas approach and the three-year AI adoption framework.
  • Companion evaluation of Eigenraum in production settings.
  • Preparatory exchanges with universities on joint application studies.

Concrete dates may shift; names of cooperating individuals and institutions are disclosed only after written consent.

Collaboration

Research collaboration

We regularly explore collaborations with universities, research institutes, funding agencies, and public authorities. For research or consortium inquiries, please get in touch.

Send research inquiry →

Frequently asked questions about the research

What does the lab research?

Three tracks with one shared core. Track A studies cognitive architectures — how a machine remembers, reasons and learns over time. Track B researches early detection of critical situations in public spaces, explicitly without biometric identification. Track C studies human-AI collaboration in knowledge work.

How do the three tracks connect?

Through the memory core. The same architecture published in Track A carries both application fields — one core, two applications. That is not a label applied afterwards; it is the reason the tracks sit in one lab at all.

Is the work openly accessible?

Yes. Publications appear on arXiv and Zenodo with a DOI, openly accessible and permanently archived; software is released under Apache 2.0. Central research documents are additionally anchored in time via OpenTimestamps, so the proof of time stays verifiable independently of the publication date.

Does the research stay theoretical?

No, and that is a deliberate commitment. The published algorithms run in our own operation, not only in papers. An architecture that does not survive operation has not passed the actual test.

What does civil liberties by design mean?

That the rights-protecting properties are architecture decisions rather than a filter added later: no biometrics, protection of helpers as an objective function instead of post-processing, and political assemblies as protected space instead of an object of study.