Gemini Memory Lifecycle Strategies

Goal

Transform transient conversation data into persistent, high-value "Memories" that allow the agent to learn about the user over time, creating a personalized experience.

The Core Lifecycle (ETL for Agents)

1. Extraction (Signal vs. Noise)

• Concept: Use an LLM to scan the raw session logs and extract only "meaningful" information, discarding pleasantries and filler.
• Method: Define "Topic Definitions" (e.g., User Preferences, Goals, Facts). If the data doesn't fit a topic, do not create a memory.
• Technique:
  • Schema-Based: Extract specific fields (e.g., {"food_preference": "vegan"}).
  • Natural Language: Extract atomic statements (e.g., "The user prefers window seats").