EXHIBIT / FROM ROOM 7 OF THE WALK
intelligent-db
A memory for AI agents that refuses unknowns instead of guessing. Every fact carries its source.
Language models fill gaps by guessing, which is the one thing a memory must never do, so I built the boring half first, a fact store that refuses to answer what it cannot prove and stamps every fact it does hold with a source and a timestamp. It also resists Sybil attacks, where one actor floods the store with fake identities to drown out the truth.
Then I red-teamed my own build, and found the pattern-completion layer was emitting deterministic noise from random weights and the associative recall was working 6% of the time. I published that finding instead of a demo.
“You built a trustworthy memory, not a mind. The valuable half is the boring, un-brain-like half: the fact store. The exciting-sounding parts are, today, set dressing.”
What it is for
- Costs today
- An agent that confidently states a fact it never actually knew, which in a regulated process is the expensive kind of wrong.
- What changes
- The agent can only cite what it can source. Ask it something it never learned and it says so, instead of inventing.
- Known in 90 days
- The rate of unsourced claims in agent output. It should trend to zero, and you can read the provenance to check.
Not yet built
Most of this does not exist yet. The markers are the point.
- The fact storeUn-inventable by construction. It refuses unknowns, and every fact carries its source and a timestamp.BUILT
- Associative recallContent-addressable memory. It works 6% of the time, and I am leaving that number in.PARTIAL
- The plannerWorking memory, a goal register, action selection, inhibition. Spec’d, nothing running yet.STUB
- Pattern completionToday it emits deterministic noise. Marked as set dressing.STUB
- Routing and attentionNot started.UNBUILT
- Salience and alarmNot started.UNBUILT
What I am building now
The planner: working memory, a goal register, action selection, and inhibition, driven by one loop.
- Sense
- Recall
- Propose
- Select / inhibit
- Act
- Record
This is how I would make an AI program audit-ready, a memory that cannot state what it cannot source and a build log that keeps its own failure numbers in the open. That just so happens to be the record-keeping and traceability duty the EU AI Act and ISO 42001 put on high-risk systems, answered at the data layer instead of in a policy document.