Memory & Knowledge Vault
Three-layer memory system automatically injected into every pipeline prompt.
GridWork HQ maintains three layers of memory for your agency. All three are injected into pipeline prompts automatically — you don't need to reference them manually.
Three layers
1. Facts
Static key-value pairs (e.g., brand.name = "Acme Agency"). Dot-notation keys. Always accurate, never inferred.
Use facts for information that doesn't change: your agency name, primary colors, team leads, client industries.
2. Rules
User-written directives with scope (global, pipeline, or client). Ordered by priority.
Examples:
- "Always respond in British English"
- "Never exceed 500 words per section"
- "For Acme Corp: use formal tone"
3. Learned
AI-observed patterns from your edit history and pipeline re-runs. Proposed automatically with a confidence score.
Example: "You prefer bullet points over numbered lists in report outputs (82% confidence)."
You accept, dismiss, or promote learned observations to rules.
Managing memory
- Memory > Facts — view and edit all facts
- Memory > Rules — view, create, and reorder rules
- Memory > Learned — view AI-proposed observations and accept/dismiss them
- Memory > Knowledge Vault — upload documents for RAG pipelines
How memory is injected
Before each pipeline run, GridWork prepends all three memory layers to the prompt. This gives your pipelines consistent context about your agency, voice, and preferences.