Agents need memory beyond the context window.
Types:
- Short-term: Current conversation context
- Long-term: Persisted facts, user preferences
- Episodic: Past interaction summaries
- Working: Scratchpad for current task
Implementation:
- Vector database for semantic search over memories
- Summarization to compress old context
- Key-value store for structured facts
Interview question: "How would you implement long-term memory?"
Store important facts in vector DB. Retrieve relevant memories based on current query. Inject into context.