Knowledge and Search

Your Knowledge is what Kojo can draw on: the knowledge built from all your meetings, plus documents. Under the hood, every meeting is processed into a knowledge graph -- see GraphRAG Memory Engine for the full technical architecture.

The Knowledge page

• Desktop: Sidebar > Knowledge (/dashboard/memory).
• Mobile: Bottom nav > Knowledge (/m/memory).

It has three tabs on desktop:

• Knowledge, entries promoted from meetings for the whole workspace to keep.
• Documents, coming soon.
• Sources and controls, which meetings feed the workspace's memory (and moving one between your personal space and the workspace), storage usage, and a summary of what Kojo can use.

Mobile shows Knowledge and Docs.

Meetings, action items, decisions and past Kojo chats each have their own page in the sidebar, so this page no longer duplicates them.

What Kojo can draw on

Kojo only ever answers from meetings and documents you have access to. A meeting scoped to a group you are not in, or restricted to specific people, never appears in your answers, your lists, or your search results. See Workspaces and Teams for how that is set.

Documents

Document uploads are coming soon. The Documents tab shows a notice until they ship.

Global search

Search across meetings, action items, decisions, and notes.

• Desktop: the search bar in the centre of the top bar.
• Mobile: the search icon in the top-right pill of the top bar.

What is indexed

Control what is searchable in Settings > Knowledge (desktop) or Settings (gear) > Knowledge Base (mobile): meeting transcripts, action items, decisions, and cross-meeting insights.

Memory window by plan

How far back Kojo can draw from when answering questions:

Meetings outside your window remain stored but are excluded from Kojo's context and search results until you upgrade.

How search works

MemorDesk uses vector similarity search: each paragraph of each meeting is converted to a 384-dimensional embedding by a locally-running AI model. Search returns results ranked by semantic meaning, not keyword overlap. You can search for concepts ("budget concerns from last sprint") and get matches even when the exact words do not appear in the transcript.

Each user's search index is partition-isolated -- queries can never cross into another user's data even under simultaneous load.