Ever wondered why your AI company brain keeps getting smarter instead of fading?
The Simple Version
Think of a company brain like a living notebook that everyone edits.
How It Actually Works
Modeling the Brain
The brain is built from about 5,000 linked markdown files, forming a wiki of interconnected knowledge.
Each file represents a piece of information, and the links create a map that the AI can traverse.
Security Through Scopes
Read and write access is defined per file, so an agent cannot silently overwrite anything.
When the agent proposes a fact, a human must approve it, and the approval carries the human’s name.
Technical Deep Dive
PromptQL extends the Hasura GraphQL engine to treat markdown files as first‑class data sources.
Queries are translated into GraphQL operations that respect the scoped permissions defined in the wiki.
Write operations are blocked unless a human explicitly signs off, and the signature is stored as metadata.
Credentials are injected at the HTTP header and SQL statement level, preventing them from residing in the sandbox.
Why This Matters
Banks and large enterprises need auditable AI systems that can’t hide mistakes.
By tying every edit to a person, organizations can trace responsibility and fix leaks quickly.
This approach also lets multiple teams collaborate on the same incident without creating separate silos.
Key Takeaways
- Company brains can leak when they are not properly scoped.
- PromptQL uses a single wiki of linked markdown files to centralize knowledge.
- Read and write permissions are enforced per file, not globally.
- Every change must be approved by a human, whose name is recorded.
- Credentials are injected at HTTP and SQL layers to stay out of the sandbox.
- The edit‑count curve is a health indicator for an AI brain.
- Scaling a brain works only when security and collaboration are built together.