Skip to main content

Workspaces & memory

The tagline promises one sovereign AI layer over everything your company knows. This page is where that layer lives.

A workspace is a living body of knowledge for a project, a customer, a team, or a workflow: the documents that define it, and — over time — the decisions, closed cases, and useful outputs that came from working it. Agents don't start from scratch; they start from the workspace. Any agent bound to it picks up the thread with the right context already in place.

Collections and bindings​

Knowledge is organized into collections — one per body of context. An agent sees a collection only if it is bound to it, at create time, in the employee YAML:

knowledge_bindings:
  - product-documents
  - matter-documents

Bindings are the scope line: an intake agent bound to the product docs cannot answer from the legal matter files, no matter what it is asked. Sharing context across agents is the same mechanism — bind two agents to one collection and they work from the same knowledge.

Documents get in by upload, by the website crawler, or by export/import when moving a collection between deployments (libraos knowledge — the dev→prod path). Retrieval over a collection is graph + vector + keyword combined, so an answer can be traced to the exact passages it used.

Grounding is visible​

You met the two fields in the quickstart: every chat response carries retrieved_chunks — the exact files the answer drew on — and a grounding verdict. Bound agents answer from your sources, and show their work. That is the property everything else here exists to serve: not that the AI sounds right, but that you can check it.

The practical side effect is cost: grounding retrieves the relevant slice of a workspace instead of re-feeding the whole history into every prompt. Shared context is cheaper than repeated context.

Memory across sessions​

Conversations are not islands. Passing conversation_id continues a thread, and the runtime's observational memory extracts durable facts from work as it happens — attributes, dates, events — so they can be recalled later without re-reading every transcript. Facts recognized as personal are routed to a dedicated per-person pool rather than the shared one; a person's private context does not leak into the team's.

The loop that compounds​

On Libra Desk this layer appears as Workspaces, and the loop closes: every case your team reviews and sends flows back into the knowledge the next draft starts from. The workspace is not a folder you fill once — it is the accumulating memory of how your company handles its work, and it is the reason the system gets better at your business the longer it runs.

All of it stays inside your deployment: the collections, the vectors, the extracted facts. On Self-Hosted, that means inside your walls — see the security model before connecting sensitive material, and Local models if even embeddings must not leave the building.