Skip to content

Memory Tool

Give Agents persistent memory to remember your preferences, learn from experience, and provide personalized service across conversations.

Why Use Memory Tool

  • Continuous service across conversations
  • Personalized responses matching your style
  • Agents improve from accumulated experience
  • Context-aware decisions based on history

Core Mechanism:

  1. User Memory: You define preferences and constraints
  2. Agent Memory: Agents record their own execution patterns and insights
  3. Categories: Both organize memories by topic
  4. Smart Search: Agents independently explore memories before executing tasks

Memory Lifecycle

Each memory has complete lifecycle tracking:

Field Purpose
ID Unique identifier (format: YYYYMMDD_xxxx)
Content Actual memory text (recommended ≤300 words)
Created At When entry was added
Updated At Last modification timestamp (auto-updates on changes)

Memory Operations

List Categories

Discover what memories exist: - Query user memories, Agent memories, or both - Returns category names and entry counts - Helps Agents understand available knowledge

Get Memories

Retrieve specific memories: - Fetch from multiple categories at once - Query across owners (e.g., user code_style + Agent learned_patterns) - Results sorted by most recently updated

Add Memories

Store new knowledge: - Add multiple entries across categories in one call - Auto-creates categories if they don't exist - Records creation timestamp

Update Memories

Refine existing memories: - Modify specific entries by ID - Auto-updates updated_at timestamp - Suitable for evolving preferences

Delete Memories

Remove outdated information: - Delete specific entries by ID - Batch delete across categories - Keep memories relevant and accurate

AI-powered memory exploration: - Agent describes what to find - Search runs in isolated subtask - Returns organized summary of relevant memories - No main conversation token overhead

Integration Patterns

Memory + Sub-Agents

Experts inherit or share memories:

Shared User Memory: All Agents in conversation access same user preferences (consistent behavior)

Independent Agent Memory: Each expert learns from their domain (code reviewer vs. writer have different insights)

Handoff Pattern: Main Agent searches memory → delegates to expert with context → expert learns from execution → updates own memory

Memory + File Tool

Combine persistent knowledge with document delivery:

Memory Guides Behavior: Agent reads documentation preferences File Tool Delivers Results: Creates api_docs.md following those preferences Learning Loop: After user approval, Agent records successful patterns in memory

Configuration Requirements

Requirements for effective memory use:

Requirement Purpose
Clear Instructions Tell Agent when to consult memory (e.g., "Always check user preferences before code generation")
Memory Access Enable memory tool in Agent configuration
Smart Search Usage Encourage use before complex tasks
Learning Prompts Instruct Agent to record successful strategies