Three identical grids. Same seed, same speed. One rule differs — only one still carries structure. AI behaves the same way: a small change to the rule, not to the compute, decides between order and collapse. Change the rule yourself: zensation.ai/en/technolog...

Technology — ZenSation Research

15 neuroscience mechanisms across a 7-layer memory architecture. Open-access preprint (arXiv:2604.23878), GDPR-native, open source.

zensation.ai

OpenAI quietly trimmed Codex's context window this week (372k → 272k). Worth noting: rented capability doesn't only disappear when a model is deprecated — it can shrink mid-use, no notice, no migration. A memory layer you own and can fully hand off doesn't get smaller by a vendor's call.

A frontier model vanished for 19 days — not a bug, an export-control order. Rented intelligence is conditional infrastructure. The layer that holds your knowledge shouldn't be. Our memory core is open, Apache-2.0, runs in your browser: huggingface.co/spaces/zensa...

Housekeeping day for ZenBrain's open core: v0.3.4 tagged as a proper release, Discussions turned on, 4 real good-first-issues posted, README now links straight to the live in-browser playground. Small stuff, but it's the difference between a repo and a project. github.com/zensation-ai/zenbrain

You can't host trust. So we didn't. ZenBrain's memory core is open source (Apache-2.0), and you can run the real code in your browser right now, no install. Live demo: huggingface.co/spaces/zensa...

'Open weights' isn't open source. And an open model doesn't make your app auditable. What must be open is the layer that touches your data: memory, orchestration, governance. Not because open is safer, but because you pick who audits it. zensation.ai/en/blog/vert...

A frontier model went dark three days after launch. Models are the perishable part. What compounds is the layer above: the memory an AI builds in use. A race for memory-portability standards is on. But interoperability isn't ownership. zensation.ai/en/blog/wenn...

What Remains When the Model Disappears?

In June, a leading model was switched off three days after launch. Models come and go. What remains is the knowledge layer above them. The real question is who owns it.

zensation.ai

A finding that reframed how I build: across 106 studies (Vaccaro et al., Nature Human Behaviour 2024), naive human+AI teams averaged worse than the better of the two alone (g=-0.23). Collective intelligence isn't a default you get by wiring two systems together. It's a design problem.

Another frontier model drops — same gap underneath: agents reason brilliantly, then forget everything between sessions. Continuity isn't a context-window problem, it's a memory-architecture problem. And the layer that remembers should be yours, not rented.

ZenBrain — a 7-layer memory architecture for autonomous agents. Working · Short-Term · Episodic · Semantic · Procedural · Core · Cross-Context. LongMemEval-500: 91.3% of long-context oracle accuracy at 1/106 of the token budget. 12/12 head-to-head wins. arXiv:2604.23878

Clarify Complexity. More Than Just AI. ZenSation — unabhängiges KI-Forschungslabor. Ein kognitiver Kern, ZenBrain, und die Spezialfelder darauf: von Wissensarbeit bis Schutz öffentlicher Räume. Civil-Liberties-Architektur als Designprinzip. Ohne biometrische Identifikation. 🔗 zensation.ai

Most "agent memory" today is high-velocity logging — log everything, retrieve some of it. That collapses two functions: scratchpad (volatile) and consolidated long-term memory (selective, schema-shaped). Brains separate them for a reason. ZenAI: 7 layers, consolidation as its own. arXiv:2604.23878

Hi Bluesky 👋 I'm Alexander, independent AI researcher in Kiel 🇩🇪 Building ZenAI — a 7-layer memory architecture inspired by neuroscience, for autonomous agents that don't forget. Open-source. GDPR-native. arXiv-published. Looking to connect with folks working on memory & agents.