Skip to main content
Knowledge is what Membrane knows about your business that no connected app can tell it: how you price, who signs off on what, what counts as urgent, the reasoning behind a decision you made last quarter. Every job, every routine, and every answer is grounded in it. Each piece of it is an article — a markdown document with a path, a title, and a one-line description of when to reach for it.

​
What an article is

Path is the article’s address in the knowledge base, written like a file path: escalations.md at the top level, policies/escalations.md inside a folder. It has to be unique in the workspace, and folders come from it — write an article at policies/escalations.md and the policies/ folder exists, with nothing separate to create. Description is the line the agent reads to decide whether to open the article at all, so write it as what this covers and when to reach for it. “When to escalate a customer issue, and to whom” gets opened at the right moment; “Escalations” does not.

​
Where articles come from

Four ways, and every article records which:
  • You write one.
  • The agent records one. When it learns something durable in the course of work, it writes it down so the next run starts with it instead of rediscovering it.
  • You upload a file. PDFs, Office documents, images, HTML, and plain markdown all become articles; the original file stays downloadable from the article it produced.
  • A connected app keeps one in sync. Synced articles are read-only mirrors: they refresh from the source and carry a link back to it, so the agent can cite where an answer came from. To change one, change it in the app it came from.

​
How the agent uses knowledge

When a session starts, the agent is handed an index of the knowledge it may use — each article’s path, title, and description, grouped by folder. It opens whole articles on demand, and searches when the index doesn’t already point at the right one. Long articles come back in parts, so nothing is out of reach. How much that index covers depends on what the session is doing. Working with you in the workspace, it’s the whole knowledge base — the first 200 articles by path, with the rest reachable by searching or listing a folder. Working on a task, it’s only the knowledge that task was given: the task names the articles and folders it runs on, and anything outside them is not listed, not readable, and not searchable for the duration.

​
Variables

Not everything an article says is prose. Where a policy comes down to a specific value — how long to wait before following up, the amount above which something needs a second approval — the article can declare that value as a variable, so the agent works from the number itself instead of reading it out of a sentence. See Variables.

​
Keeping knowledge trustworthy

History. Every write that changes what an article says — its title, its body, its variable declarations, or a variable’s value — is recorded, along with who made it: a person, an agent session, a task, or a sync. When an edit was based on a task, workspace element, thread link, or outside URL, those sources appear under Based on in the article’s attribution and history. If a specific message prompted it, the note there says which message and why. A re-path or a retag isn’t a change to what the article says, so it doesn’t produce one. Confirming. “Nobody has changed this in a year” and “somebody checked this last week” are different facts, and only the second one lets a reader trust an old article. When you’ve checked an article against reality and found nothing to change, confirm it: that records who checked and when, without touching the text. Archiving. Deleting an article removes it from the knowledge base and hides it from the agent, but keeps it.

​
Moving and renaming

Changing an article’s path moves or renames it. Change a folder’s path and everything beneath it moves at once. Synced articles stay where their sync puts them.

​
Exporting and importing

To back up a knowledge base, copy it into another workspace, or keep it in version control, export it — all of it, or one folder — as a portable list and import it back. Import matches on path, so it creates what’s new and updates what already exists; paths owned by a sync are left alone.

​
Who can see knowledge

Knowledge grounds the workspace as a whole: everyone working in it reads and writes the same knowledge base, and there is nothing to share article by article. Customers embedded in your own product — tenants using Membrane through your app — have no access to it.