# Fact

A **fact** is a single grounded statement Connect holds — *the returns window is fourteen days* — extracted from a [knowledge](/docs/glossary/knowledge/) source or from a conversation, and carrying where it came from. Facts are managed one at a time, which is what makes a wrong one correctable without editing a document.

- **Status:** Available
- **Audience:** both
- **In the app:** #/knowledge
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/glossary/fact/

## One statement, one origin

The unit matters. A source is a document and can be long, mixed and partly out of date; a fact is one sentence that is either right or wrong. Splitting them gives you a correction that takes effect immediately, without waiting for somebody to reissue a PDF.

Provenance travels with the statement, which is the other half of the point. A fact you can trace answers the question that follows every surprising reply: *where did it get that?* Without provenance the only available response is to distrust the whole store, and a store nobody trusts stops being used.

**Extracted** — Pulled from a source you added. The source is the origin, and re-reading it later is possible.
**Observed** — Noticed in a conversation and kept as a statement. The conversation is the origin.
**Corrected** — Edited by a person. The correction is what is asserted from then on.

## How facts reach an answer

A handful at a time. Only a few facts accompany any single piece of work — on a live call the budget is roughly four — because everything given to a model before it speaks delays the first word. That constraint produces a counter-intuitive rule for maintaining the store: **a smaller set of sharp facts beats a larger set of vague ones**, because selection, not capacity, is the scarce thing.

It also means near-duplicates are actively harmful. Two facts saying almost the same thing with different numbers do not average out; one of them gets selected, and which one is not something you control from the screen. When you find a pair like that, keep one.

## What a fact is not

**A knowledge source** — The document is the source; the sentence pulled out of it is the fact. Deleting the source does not silently rewrite the statements taken from it, which is why both screens exist.
**A [memory](/docs/glossary/memory/)** — Memory is about this business and these people and is not for quoting outward. A fact is a statement about the business that may be repeated to a customer. The test is whether you would be content to see it in a reply.
**Something the model believes** — A model's general impression of your industry is not a fact in this sense and carries no provenance. The whole apparatus exists to prefer the traceable statement over the plausible one.
**A guarantee that it is true** — A fact is what the workspace asserts, not what reality confirms. An out-of-date fact is asserted confidently until someone corrects it — which is why review matters more here than in a store nobody quotes from.

One consequence for support: when a reply contains a wrong number, the fix is usually one fact rather than a prompt, a rule or a setting. Correcting it there changes every future answer that would have used it, on every channel, without touching how Connect behaves in any other respect.

## Questions

### How do I correct a fact Connect keeps repeating?

Edit the statement itself on the facts screen rather than arguing with it in conversation. A correction made only in a reply is a decision about that reply; the stored statement is what the next answer uses.

### If I delete a knowledge source, do its facts disappear?

They are managed separately, so review both. That separation is deliberate — it stops a document tidy-up from silently changing what Connect asserts — but it does mean a decommissioned source can leave statements behind.

### Can Connect invent a fact?

It can generate a plausible sentence, which is why the store keeps provenance and why grounded answers are preferred to recalled ones. A statement without an origin is the thing to be suspicious of.

## Related

- [Knowledge](https://connectbyjbrh.com/docs/glossary/knowledge/)
- [Memory](https://connectbyjbrh.com/docs/glossary/memory/)
- [Knowledge in Connect](https://connectbyjbrh.com/docs/knowledge/)
- [Grounding an answer without pretending to have a vector database](https://connectbyjbrh.com/research/grounding-without-a-vector-database/)
- [Why uncertainty is a valid answer from a business agent](https://connectbyjbrh.com/research/uncertainty-is-an-answer/)

## What this page is based on

- `docs-source/facts.py` — CANONICAL_TERMS Fact; MEASURED knowledge_budget_facts
- docs-source/sources/GENERAL.md §4 — knowledge.py, the facts editor
- Connect capability registry (docs-source/facts.py) — facts_editor: available
