Guide: Mastering Memory Banks

Memory Banks are powerful, but they work best when populated with high-quality data. This guide explains how to curate your banks for maximum accuracy.

Examples, Not Definitions

The most common mistake is providing definitions. Memory Banks work by semantic similarity. You should provide varied examples of how a user might ask for something.

Bad:

{ "query": "Definition of reset password", "correct_tool": "reset_password" }

Good:

{ "query": "I forgot my password", "correct_tool": "reset_password" }
{ "query": "Change my login code", "correct_tool": "reset_password" }
{ "query": "Cant login help", "correct_tool": "reset_password" }

Diversity is Key

Try to cover different phrasings, slang, and levels of politeness.

• “Give me the report” (Direct)
• “Could you please download the Q3 summary?” (Polite)
• “where report” (Lazy/Chat)

Bank Granularity

Should you have one big bank or many small ones?

• Core Bank: Create a “Global” bank for common commands (Login, Logout, Help) and assign it to all Apps.
• Feature Banks: Create specific banks for specific features (e.g., “Reporting”, “User Management”). Assign them only to Apps that have those features enabled.

Conflict Resolution

If two banks contain similar queries mapping to different tools, the Priority setting in the App->Bank assignment controls which one takes precedence.