memory-management

Memory Management Pattern

Overview

The Memory Management Pattern enables agents to retain and utilize information from past interactions, observations, and learning experiences. Without memory, every agent turn is isolated — the agent has no awareness of prior context. With effective memory management, agents can maintain conversation continuity, personalize responses, track task progress, and improve over time.

Core Principle: Give agents the ability to remember — short-term for the current interaction, long-term for knowledge that persists across sessions.

When This Skill Applies

Activate this pattern when:

• Agents must maintain context across multiple conversation turns
• User preferences or past behaviors should influence future responses
• Multi-step tasks require tracking progress and intermediate results
• Agents need to recall information from previous sessions
• Personalization and continuity are key user experience requirements
• RAG (Retrieval-Augmented Generation) is needed for knowledge-grounded responses

Rule of thumb: If an agent needs to "remember" anything beyond the current prompt — use Memory Management.

Memory Types

Short-Term Memory (Contextual Memory)

• Lives within the context window of the current LLM call
• Contains: recent messages, tool outputs, agent reflections, session state
• Limited capacity: context windows have token limits
• Ephemeral: lost when the session ends
• Management strategies: summarize older segments, prioritize key information, use long-context models

Long-Term Memory (Persistent Memory)

• Stored outside the agent's immediate context in external systems
• Storage options: databases, knowledge graphs, vector databases
• Semantic search: vector embeddings enable similarity-based retrieval
• Persistent: survives session terminations, restarts, and time gaps
• Use cases: user preferences, learned knowledge, historical records

DEFINE → PLAN → ACTION Workflow