Knowledge

Knowledge is the searchable memory of your workspace. It is how an agent in a huddle can find the decision made two months ago instead of re-deriving it.

Knowledge is not the company profile

Both are things agents read, and they play different roles.

	Company profile	Knowledge
What it is	The canonical description of your company	Everything the workspace has produced
How it changes	Only with your approval	Automatically, as work happens
Shape	Structured fields	Passages, searched by meaning
Authority	Ground truth	Evidence, quoted with its source

Put simply, the company profile says what is true about you. Knowledge says what has been said and done.

What gets ingested

As work happens, the platform indexes it. The corpus holds:

• Huddle messages and huddle summaries
• Tasks, with their specs
• Workspace items, the documents and notes kept in the workspace
• Research findings from research jobs
• Content assets produced by marketing work
• Artifacts created in huddles and runs
• Repository files, for repositories that have been indexed

Ingestion follows the work rather than a schedule: when a huddle summary is generated or a workspace item is updated, that change is what triggers the indexing, and deleting the source removes it from the corpus.

Everything is scoped to your workspace, and a search never crosses into another workspace’s corpus. Inside your workspace the corpus is organised by project, which is what gives a search its reach. See how far a search reaches below.

knowledge_search

knowledge_search is the tool agents use to query the corpus. It takes a question in plain language and returns the passages that answer it, with what they came from.

The same tool appears on every surface where an agent thinks, and it is the same contract each time: in huddles, in conversations with Eva, and inside coding runs, where it is exposed to the coding agent through a small local shim. See knowledge in runs. The reach rule below is the same on all of them.

Agents are pointed at it for a specific purpose: check whether something already exists before rebuilding it. A search returns up to twenty passages, eight by default, ranked by relevance rather than recency.

How far a search reaches

An agent searches the project it is working in. A huddle held in one project searches that project, a coding run searches its task’s project, and a conversation with Eva about a project searches that project.

That default is deliberate. “What is our positioning?” means one thing in a huddle about a single product and something else in a huddle about the whole company, and the two questions are word for word identical. Searching too narrowly corrects itself in one turn, because you say so and the agent looks wider. Searching too widely is invisible: the answer reads as though it all came from here.

So reaching further is always something the agent does on purpose:

• Named projects. An agent can search one or more projects you name instead of the current one. In a huddle it can look up the projects in your workspace first, so it does not have to guess which one you meant.
• Every project. An agent can search the whole workspace when the question is genuinely about more than one product. In a huddle it judges that from the huddle’s title and objective, which is why both are worth writing properly.
• Nothing to narrow to. A conversation with Eva that belongs to no project has no project to search, so the search covers the whole workspace and the result says that is what happened.

Anything found outside the current project is labelled with that project’s name, on every surface, so a passage about one product cannot be quoted as though it came from another. Agents are told to attribute it that way when they use it.

Next steps

• Knowledge search: searching the corpus yourself.
• Knowledge in runs: how a coding agent reaches the corpus.
• The company profile: the canonical context that sits beside it.