Blind Spot.

Governed memory for AI coding agents — provenance, scope, rollback, and human promotion, not just a bigger vector store.

Mori tackles a real, growing problem: agents need memory, but ungoverned memory contaminates work across projects. It appears unusually built-out relative to its public visibility — support across multiple coding-agent environments, and a clear theory of why shared memory fails.

This may not just be a developer tool. If the governed-memory model works, the same pattern could apply to law firms, research groups, and design studios — anywhere AI agents need institutional memory without carrying the wrong context into the wrong task.

Structured relationships pulled from messy cultural text, turned into an explorable map of French food.

This isn’t a clean “nobody noticed” case — the technical community did take note. But people saw a clever graph of restaurants and may have missed the builder signal: Cantelobre took a messy cultural corpus, extracted structured relationships from unstructured writing, compared models, controlled costs, handled hallucination, cleaned entities, and shipped an explorable interface.

It may have been underpriced because it looked like a charming culinary side project rather than a repeatable applied-AI method. The same workflow could scale to film crews, galleries, hospitals, or academic labs — any field where hidden relationship networks are buried inside public text.