Knowledge

The knowledge stack gives the AI assistant context about your campaigns, decisions, and reference material. Each layer has its own tab on the Knowledge page.

Open Knowledge in the sidebar, under the Workspace group. Besides the three tabs below, the page has an Overview of every retrieval source and an Agent memory tab for the key-value memory the agent writes while it runs.

The Three Knowledge Tiers

Briefing — Project Memory

Navigate to Knowledge > Briefing.

Project memory is a free-form document that is injected into every AI prompt. Use it to record:

• Preferred bidding strategies and budget norms
• Brand voice rules and audience exclusion preferences
• Reporting cadence and team conventions
• Campaign naming patterns

Edit the memory directly in the text editor and click Save. Changes take effect on the next conversation.

Notes — Knowledge Records

Navigate to Knowledge > Notes.

Knowledge records are structured entries extracted from past conversations and decisions. The AI creates them automatically as you work — you don't need to write them manually.

Use the search and filter controls to find records by:

• Topic — what the record is about
• Campaign — which campaign it relates to
• Date — when it was created

Records give the AI historical context when proposing new actions. For example, if you reallocated a budget last quarter, the record of why helps the AI make better recommendations this quarter.

Reference Docs — Document Search

Navigate to Knowledge > Reference docs.

Knowledge bases are collections of uploaded documents that the AI can search using vector similarity (semantic search). Use them for:

• Brand guidelines and style guides
• Compliance policies and legal requirements
• Campaign playbooks and standard operating procedures
• Market research and competitive analysis

Creating a Knowledge Base

1. Click New Knowledge Base
2. Enter a name and optional description
3. Add text-based content — plain text, Markdown, CSV, or JSON
4. The platform chunks, embeds, and indexes the content automatically

Processing happens in the background — large documents may take a few minutes to become searchable.

How Document Search Works

When the AI needs specific reference material, it searches your documents by semantic similarity using pgvector. Relevant passages are automatically injected into the prompt context alongside project memory and knowledge records.

How the Tiers Work Together

For any given prompt, the AI assembles context in this order:

1. Always: Project memory (Briefing) is injected in full
2. On demand: Knowledge records are searched and the most relevant results are included
3. On demand: Document passages are retrieved via vector search and included

This layered approach keeps the context window focused — not every document is loaded every time, only what's relevant to the current question.