Nova Remembers What Your Team Teaches It
Corrections become permanent, verified, org-wide knowledge — with the evidence behind every fact. Nobody ever has to teach Nova the same thing twice.
Ask Nova, get SQL + charts
Meet Nova
One key per partner. No credentials shared.
Build an API
+more
Scheduled sync to 6+ platforms
Explore
+more
Query APIs with SQL, no code
Explore
+more
Claude, Cursor, ChatGPT & Grok talk to your data
Connect AIEvery team carries hard-won knowledge about its data that lives nowhere in the schema: that receipts table is dead — current data lives in shipments. “Closed” includes cancelled orders. That back-reference is only 10% populated, don’t trust it for joins. Traditionally you re-explain these things to every new analyst — and to every new AI conversation.
Taught Knowledge ends that. When someone corrects Nova or teaches it something durable about your data, Nova can keep it — permanently, for the whole organization. But unlike a chatbot “memory” that stores whatever it heard, Nova refuses to remember anything it can’t prove. This guide explains the whole system.
The Core Principle: Nothing Is Remembered Without Proof
Every learned fact passes through one of two verification doors before it takes effect. There is no third path — a model inference that merely “sounds right” is rejected by design.
Door 1: The SQL Probe
Schema facts — what a table is, what a column means, whether a table is deprecated — are verified by a small read-only query against your actual database, through your own Network Agent. The probe SQL and its result are stored as evidence, re-runnable forever.
Door 2: Human Confirmation
Business rules and metric definitions can’t be proven by a query — they activate only when a person confirms them. Explicitly (“yes, that’s right”), or behaviorally: Nova applies the pending rule in a query, it verifies, and you accept the results. You watched it work.
How a Fact Becomes Knowledge
Facts are deliberately compact — a few hundred characters each. If you teach Nova something with several independent parts (a definition, a join recipe, and a counting convention), it splits them into separate facts and verifies each one individually. Small, atomic facts are easier to verify, correct, and retire than paragraphs.
What Nova Can Learn
| Fact type | Example | Verified by |
|---|---|---|
| Deprecated table | “That receipts table is dead — use shipments” | A range-coverage probe: the replacement must fully cover the old table’s data before it’s excluded. If the old table holds history the replacement doesn’t, it becomes a historical pointer instead — past data is never orphaned. |
| Table role | “This is a view over the orders table” | Catalog probe — the table exists and is queryable |
| Column meaning | “Status codes 18/14/16/20 all mean open” | The column exists and the claimed values actually appear in the data |
| Join reliability | “That back-reference is only ~10% populated” | The measurement is the fact — Nova runs the population query and stores the result |
| Business rule / metric | “Closed means Invoiced or Cancelled” | Human confirmation — explicit or behavioral |
A Fact’s Life Cycle
Staged, awaiting verification. Completely inert — zero influence on Nova until it passes. Unconfirmed proposals auto-retire after 30 days so nothing piles up.
Verified and serving Nova now, org-wide. Every use is tracked; every success and contradiction is counted.
Reality disagreed. The fact is suspended from Nova’s reasoning immediately and its original evidence probe re-runs. Pass → reinstated, flagged for review. Fail → retired.
Removed from every part of Nova’s reasoning — kept for the record with its full history and evidence.
Tiers: Learned Knowledge Makes Nova Cheaper, Not Heavier
Naive AI memory grows without bound and drowns the model in notes. Taught Knowledge inverts that — most learned facts reduce what Nova has to read:
Tier 0 — Exclusion
A verified deprecated table is removed from Nova’s working context, leaving a one-line pointer to its replacement. Learning that actually saves tokens.
Tier 1 — Description
Table and column facts replace Nova’s AI-generated guesses in place. Zero added cost — and a description hand-written by your team always outranks a learned one.
Tier 2 — Injected
Your most-used, best-corroborated business rules ride along in every conversation, under a strict size budget that can never blow up.
Tier 3 — On-Demand
Everything else is searchable the moment it’s relevant, at zero standing cost. Long-tail rules and procedures live here.
The Taught Knowledge Tab
Everything Nova has learned lives under Nova → Schema Intelligence → Taught Knowledge (admins and creators):
Facts
Browse every fact with filters. Expand any row for its evidence — the exact probe SQL and result, or the confirmation — plus full version history.
Business Rules
Rules and metric definitions, split by injected vs on-demand, disputed flagged red, pending proposals showing confirmation progress.
Settings
Org-wide switches — learning on/off, member teaching, exclusions, rule capture — plus platform guardrails shown read-only for transparency.
Teaching Roster
Per-person “can teach Nova” toggles. Block a noisy teacher without turning learning off — blocked users still benefit from everything the org has taught.
Activity
A timeline of everything — who taught, revised, disputed, or retired what, and when.
Taught facts also surface where your eyes already are: the Schema Intelligence tree shows Taught badges on tables and columns with learned knowledge, deprecation chips on excluded tables (dimmed but always visible — never hidden), and a per-database knowledge summary.
Provenance: Every Fact Shows Its Work
Every fact permanently records who taught it (with their role and photo), when, in which conversation, and the evidence that verified it. Ask Nova directly — “who taught you that closed includes cancelled?” — and it answers with the real provenance. Admins can revise a fact (revisions re-enter verification — editing authority never bypasses proof) or retire it, from the UI or right in chat.
Self-Maintenance: Knowledge That Stays True
Usage telemetry
Every fact tracks how often it’s used, succeeds, and gets contradicted.
Contradiction handling
Two strikes and a fact suspends itself pending re-verification against live data.
Decay & promotion
Unused rules demote out of the injected set and archive; rules that keep proving useful earn promotion.
Exclusion self-repair
Every schema rescan re-checks excluded tables. Data reappears in a “dead” table? The exclusion suspends automatically. An exclusion can never outlive its proof.
Measurement refresh
Measured facts — like join reliability percentages — re-run on schedule as your data drifts.
Safety and Control
| Guarantee | How it’s enforced |
|---|---|
| Strictly org-scoped | Your facts are yours. Nothing learned in your organization is ever visible to, or influenced by, any other organization. |
| Read-only verification | Probes run through your own Network Agent with the same read-only enforcement as every Query Streams query, under hard time and row limits. |
| Escape hatches everywhere | An excluded table is a routing preference, not a wall — it can always be queried explicitly, and SQL referencing it gets a gentle redirect notice, never an error. |
| Guardrails | Per-conversation proposal limits, per-organization fact ceilings, and compact per-fact size caps prevent runaway growth or prompt-injection abuse. |
| On by default, off in one switch | Works out of the box with nothing to configure. Turn any part off in Settings — nothing is deleted, and flipping back on restores everything instantly. |
Frequently Asked Questions
Why didn’t Nova remember what I taught it?
Can a wrong fact poison Nova for everyone?
Does learning slow Nova down or cost more?
Who can teach, and who can manage what’s been taught?
What happens if we turn learning off?
Teach Nova Once. Benefit Forever.
Connect your data source, ask questions in plain English, and correct Nova like you’d correct a new analyst — it keeps what it verifies. Free to start, no credit card required.
Related guides: How Nova AI Learns Your Database Schema Automatically | Text to SQL with Nova AI | AI Powered Analytics
English 
Spanish 
German 
French 
Portuguese 
Japanese 
Korean 
Chinese 
Italian 
Dutch 
Indonesian 
Polish