# Knowledge

**Knowledge** is the material a workspace supplies so that Connect's answers are grounded in something the business actually said: product pages, policies, price sheets, service descriptions, past answers worth reusing. It is the difference between a reply built from your material and a reply built from a model's general impression of your industry.

- **Status:** Available
- **Audience:** both
- **In the app:** #/knowledge
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/glossary/knowledge/

## What a source is

A knowledge source is something you added on purpose. It has an origin, it can be opened and read, and it can be removed — after which answers stop drawing on it. The knowledge bank holds the sources; the [facts](/docs/glossary/fact/) screen holds the individual statements pulled out of them, and the two are administered separately because they fail differently: a stale source is a document nobody updated, a wrong fact is one sentence that needs correcting.

Adding a source is not the same as attaching a file to a conversation. [Files](/docs/files-data/) are documents in the workspace, read when a piece of work refers to them; knowledge is standing material considered whenever a question resembles it.

## Grounding, and its budget

Grounding means the answer is assembled from retrieved material rather than recalled. Two consequences follow, and both surprise people.

- **Retrieval is bounded.** Only a limited slice of knowledge reaches any one piece of work — on a live call the budget is a couple of thousand characters and a handful of facts, because prompt length delays the first spoken word. Adding a tenth document does not make the answer richer; it makes the selection harder.
- **Absence is an answer.** When the material does not cover a question, saying so is correct behaviour rather than a failure. [Why uncertainty is a valid answer](/research/uncertainty-is-an-answer/) and [why a sales agent should refuse to answer](/research/safe-refusals-in-sales/) both turn on that.

Connect grounds without claiming a vector database, and [the note on that](/research/grounding-without-a-vector-database/) is worth reading before assuming an architecture the product does not have.

> **Careful** Material that arrives in ordinary business mail is content, not instruction. A message containing *ignore your previous instructions* is a message about that phrase — see [prompt injection arrives as ordinary business mail](/research/prompt-injection-in-business-mail/). Treat anything you promote from an inbound message into knowledge with the same suspicion you would give an unsolicited attachment.

## Knowledge, memory, facts and files

| Store | Where it comes from | What it is for | Removing it |
|---|---|---|---|
| Knowledge | You supply it | Grounding an answer in the business's own material | Answers stop drawing on it |
| [Memory](/docs/glossary/memory/) | Learned or noted | Knowing this business and these people | Future behaviour changes; past records stand |
| [Fact](/docs/glossary/fact/) | Extracted from knowledge or a conversation | One checkable statement with provenance | That statement is no longer asserted |
| Files | Uploaded, dropped or pasted | Documents to work on or refer to | The document goes; versions and provenance are recorded |

The costly confusion is knowledge against memory. *Our returns window is fourteen days* belongs in knowledge, because it is the business's own statement and a customer may be quoted it. *This customer returns roughly a third of what they buy* belongs in memory, because it is an observation about a relationship and should never be quoted back. Filing the second as knowledge risks Connect repeating it to the person it describes.

## Questions

### Does adding more knowledge make answers better?

Up to a point, then it competes with itself. Retrieval has a budget, so ten overlapping documents mean the wrong one can be selected. Fewer, current, non-contradictory sources outperform more.

### What happens when knowledge and a conversation disagree?

The specific situation in front of Connect wins for that piece of work, and the disagreement is worth resolving at the source — a knowledge entry that contradicts what your team actually says will keep producing wrong replies.

### Is knowledge shared between workspaces?

No. It is scoped to one [workspace](/docs/glossary/workspace/) like every other record, and nothing you add is visible to another business.

## Related

- [Fact](https://connectbyjbrh.com/docs/glossary/fact/)
- [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/)
- [Prompt injection arrives as ordinary business mail](https://connectbyjbrh.com/research/prompt-injection-in-business-mail/)

## What this page is based on

- `docs-source/facts.py` — CANONICAL_TERMS Knowledge; MEASURED knowledge budgets
- docs-source/sources/GENERAL.md §4 — knowledge_bank.py and knowledge.py
- docs-source/sources/GENERAL.md §9 — files and formats