AI Memory

AI Memory gives you a way to fine-tune how the GoodData AI Assistant understands and answers business questions. It lets you define additional workspace knowledge that guides the Assistant’s tone, terminology, and logic so users always get accurate and consistent responses.

Organization Admins or Workspace Admins with the AI_ASSISTANT permission can access this feature from Catalog → AI Memory.

Example: If users often ask for “sales performance,” but your semantic model does not include a metric with that exact name, you can create an instruction that tells the AI to use Total Sales and Net Sales together whenever someone asks about “sales performance.”

This ensures consistent, predictable behavior without modifying the underlying data model.

How AI Memory Works

You can use AI Memory to give the AI Assistant Instructions. These are rule-based mappings that define how the AI should respond to specific types of questions. These can be simple mappings or conditional rules that apply only when relevant.

Examples

• If the user asks about sales performance, show Total Sales and Net Sales.
• If the user asks about regional performance, filter by Europe and North America.
• Use a professional tone in all responses.

Instructions can control both content (what information is returned) and style (how responses are written).

Each instruction includes:

• Short name: A clear identifier for easy management.
• Memory content: The text of the rule or behavior.
• Application mode: Choose whether the rule applies “always” or “when relevant.”

Suggested AI Memories

Suggested AI Memories help you improve AI Assistant responses based on user feedback. When users give negative feedback on an answer, GoodData can generate suggested memory items that you can review and enable.

Suggestions appear as disabled memory items with a name that starts with Suggestion:. Disabled suggestions do not change AI Assistant behavior.

How Suggestions Are Created

1. A user asks the AI Assistant a question in a workspace.
2. The AI Assistant answers.
3. The user gives negative feedback if the answer is incorrect or misleading.
4. GoodData automatically evaluates feedback and creates suggested memory items.
5. An admin reviews the suggestions in Analytics Catalog → AI Memory and decides what to enable.

Example

A user asks: What was our best-selling product last month?

If the AI Assistant answers using the wrong definition, the user can give negative feedback. Later, the system may suggest memories such as:

• Suggestion: When users ask about best-selling product, prefer Units Sold over Revenue.
• Suggestion: Interpret last month as the previous calendar month.

You can enable the suggestions that match your definitions. After that, the AI Assistant uses the enabled memories for similar questions in the same workspace.

What You Might See

• No New Suggestions
  After a feedback run, the suggestions list can remain empty.

• Similar Suggestions
  Multiple feedback events can produce overlapping suggestions. Enable the best one and keep the rest disabled, merge them into one, or dismiss them all.

Current Limitations

• AI Memory entries are static and must be updated manually.
• Too many overlapping entries may reduce AI flexibility.
• Entries apply globally within a workspace and may cause conflicts in multi-team environments.
• Fallback visualization settings are not stored.

Frequently Asked Questions

Q: Who can define AI Memory entries? A: Organization Admins or Workspace Admins with the AI_ASSISTANT permission.

Q: How is AI Memory different from data modeling? A: Data modeling defines structure, while AI Memory defines interpretation and expected behavior.

Q: Will AI Memory override semantic layer rules? A: No. It supplements the semantic layer but cannot override logic or permissions.