An independent research lab in Kiel, Germany. We research collective intelligence on cognitive architecture — humans and machines together, on a structure that retains knowledge.
Every number on this page is backed by a public record, a passing test suite, or a reproducible build. Nothing here asks to be believed.
- On LongMemEval-500 (500 questions), with a measured ZenBrain arm that is a Python port of a scoring class from our test harness (not the seven-layer system, not an import of the packages), 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 (3 competitors × 3 LLM judges, Bonferroni-corrected, version-matched). Input tokens are counted in the answer step only and building the memory is not counted. The paper also prints where ZenBrain loses: retrieval-proper metrics to a competing system, an aggregate-F1 metric to lexical search. Source: paper version 9, Zenodo 10.5281/zenodo.23832288, also arXiv:2604.23878v4.
- The memory library is extracted from ZenAI, the AI platform we ran in our own operation (440K+ lines of TypeScript and 12,000+ test cases in its public snapshot of 8 May 2026) — not a research toy. Releases are built in public CI, and the current version of every package is published to npm with build provenance, so it traces back to its commit.
- The research trail is fully public: open records with DOIs, an arXiv preprint, and defensive publications establishing prior art.
🧠 ZenBrain — dynamic agent memory
7 layers, FSRS-style spaced repetition, Hebbian learning, emotional tagging, sleep consolidation. Pure TypeScript, no third-party dependencies in algorithms and core, tested in CI on Node 22, 24 and 26.
npm install @zensation/algorithms # the library
claude plugin install zenbrain --marketplace zensation-ai/zenbrain # memory for Claude Code- Try it in your browser — runs the published code, no install.
- Packages:
@zensation/algorithms·@zensation/core·@zensation/adapter-postgres·@zensation/adapter-sqlite - Drop it into what you already use:
@zensation/mcpgives any Model Context Protocol client five memory tools over one local SQLite file, reaching the working, episodic, semantic, procedural and core layers;@zensation/ai-sdkis Vercel AI SDK middleware that recalls before the model call and stores the turn after it. - New here? Pick a good first issue or ask in Discussions.
🗳️ voteworth — independence-weighted aggregation for LLM voter and judge panels
Measure how often two voters of an LLM judge panel or voter ensemble pick the same answer when both break from the plurality, report Kish's effective number of votes by raw agreement, and correct the vote for voters that do, naming the panel shape where it cannot. Closed-form, label-free, numpy only.
pip install voteworth🤖 ZenAI — the platform it came from
The self-hosted AI OS ZenBrain was extracted from: 60 AI tools, multi-agent system, knowledge graph, real-time voice. Published as a frozen, readable snapshot (8 May 2026). Its terminal agent @zensation/cli is retired and deprecated on npm since 1 October 2026; the maintained memory tooling is @zensation/mcp.
- Preprint: arXiv:2604.23878 — ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems
- Corpus: Open records on Zenodo — preprints, software / reproduction packages, datasets and notes, every preprint under CC BY 4.0, each with its own version and concept DOI
- Author ID: ORCID 0009-0001-1793-012X
- Prior art: defensive publications on Technical Disclosure Commons
- Benchmarks are reported with the protocol, the effect sizes and the cases where the system loses — including the metric on which a competing system wins.
- Library questions → GitHub Discussions
- Research, replication, consortium inquiries → research@zensation.ai
- Everything else → open-source@zensation.ai · zensation.ai
Company: ZenSation Enterprise Solutions · Kiel, Germany · Code: Apache 2.0 · Records: CC BY 4.0 / Apache 2.0