Design Agent Memory Architecture for Long-Running Tasks

Design a memory architecture for an AI agent that executes multi-hour tasks involving hundreds of tool calls. The agent needs to remember what it has done, why it made decisions, what it learned, and user preferences — across both a single session and multiple sessions spanning days.

Scale Requirements

• Single task can span 8+ hours and involve 500+ tool calls
• Agent must maintain context across 50+ concurrent user sessions
• Context window for the underlying LLM is limited to 200K tokens
• Long-term memory should persist indefinitely and span thousands of past interactions
• Retrieval latency for relevant memories: under 100ms

Design Requirements

1. Define a taxonomy of memory types the agent needs and their distinct requirements.
2. Design the working memory system that keeps the agent on track within a single session.
3. Design the long-term memory system for cross-session recall.
4. Explain how you handle memory conflicts (contradictory information from different sessions).
5. Address privacy and security: what should be forgotten, and when?