Client Context Ingestion (Onboarding)

Client Context Ingestion (Onboarding)

1. 01

   Inventory every client artifact that currently lives outside the memory layer

   Brand guides, campaign histories, audience segments, and architecture decisions are the four categories that most often sit in shared drives or individual inboxes. List each artifact with its owner and last-modified date before any connector is wired.

2. 02

   Assign a canonical project namespace and tag taxonomy before writing the first memory entry

   Knownbase organizes notes by project and tag, so a namespace collision at onboarding becomes a retrieval failure six weeks later. Agree on naming conventions with the delivery lead and lock them.

3. 03

   Connect the memory server to each agent in the stack through its MCP endpoint

   Bourdon federates memory across Claude, Codex, Cursor, Copilot, and Devin so a fact learned by one agent is recognized by the others. Confirm each connector authenticates against the same namespace.

4. 04

   Seed the graph with entity relationships, not just flat documents

   Cogni retrieves connected facts through entity-graph spreading activation, which reaches multi-hop relationships a vector store cannot. Write client entities (brand, product line, stakeholder, campaign) as linked nodes from day one.

5. 05

   Run a retrieval test with three real client questions the team asks weekly

   Ask the memory layer what the client's tone guidelines are, which campaign underperformed last quarter, and who approved the current creative direction. If any answer requires a re-prompt, the ingestion is incomplete.

6. 06

   Document the export path and verify it produces a portable file

   Vendor lock-in is the category's structural risk: if the memory layer does not export cleanly, institutional knowledge leaves with the tool. Test the export before the client's first deliverable ships.

7. 07

   Set a 30-day review to prune stale entries and confirm freshness signals

   OzBrain routes new knowledge to the appropriate article and tracks freshness, but stale entries degrade retrieval quality over time. Schedule the review as a recurring calendar hold, not a one-off task.