How do you implement memory systems for long-running AI agents?

AI agent memory enables context persistence across conversations and tasks. Three memory types: short-term (current conversation buffer, typically last 10-20 messages), working memory (task-specific context maintained during multi-step operations), and long-term (persistent knowledge accumulated over time). Short-term implementation: sliding window of recent messages with summarization of older context. Working memory: maintain a structured state object tracking current task goals, completed steps, intermediate results, and pending actions. Long-term memory implementation: store important facts, user preferences, and learned patterns in a vector database with metadata (timestamp, source, confidence). Retrieval: on each interaction, embed the current context and retrieve relevant memories from long-term storage. Memory management: implement decay (reduce relevance of old memories), consolidation (merge related memories), and forgetting (remove contradicted information). Frameworks: LangChain MemoryModule, MemGPT (manages memory like an OS), and Zep (purpose-built memory server). Architecture for production: use Redis for working memory (fast, ephemeral), PostgreSQL for structured long-term storage, and pgvector for semantic memory retrieval. Key challenge: determining what to remember versus forget, and when to surface memories without overwhelming the context window.