A lab with a product, not a product with a lab

Mycelic is the name of the research programme and of the lineage-first fabric; NeuralGraph is the engine.12 The engine is built by Ali Novruzov as part of ongoing work on persistent memory, agent orchestration and graph-based intelligence.3 Track contributors are credited to the lab, not by name, while the manuscript is under review; individual credits will be added after the review decision.4

The legal entity and year in the footer are set by the owner. No customer is named anywhere on this site because none is named in the sources.

From the engine’s README, verbatim:5

1. Memory should preserve structure, not just text.
2. Retrieval should use multiple signals.
3. Time and speaker identity are first-class metadata.
4. Answers should remain grounded in traceable evidence.
5. The system should run locally when possible.
6. Every retrieval failure should be measurable.

And from the fabric: capability coherence without state convergence: independent memory holders jointly reconstruct an answer no single one of them holds, without merging into one global graph.2

Benchmark first, then architecture. We separate evidence classes and refuse to pool them.6 We publish the conditions where the approach loses next to the ones where it wins.7 We audit our own harnesses: the current LoCoMo numbers exist because an audit found two leakage channels in the earlier one and removed them.8 Our organisational benchmark gives centralised baselines strictly more information than our own system and is built so that Mycelic can lose.910 Every number on this site has a row in a register that wins if a page disagrees.

Claims register →

• Graph memory, persistent memory, reconstruction or hierarchy as standalone inventions.
• Deployment of the complete architecture anywhere; no experiment validates it end to end.11
• That privacy and authorization are experimentally established guarantees.12
• That the engine is open source; the repository is public to read, but no licence has been selected.1319
• Accuracy comparisons with other systems; judge choice alone moves the same answers by 51 points.14
• Any latency or hardware specification.15
• Enterprise readiness, certification, compliance or an SLA.16

• Research partners who want to reproduce the deterministic fixture, run the LoCoMo harness under their own judge, or plug a measured model profile into the organisational benchmark.1718
• Reviewers who need artifact access.
• Organisations that want an evaluation on their own hardware.

One contact path.