agent-memory-systemslisted

# Agent Memory Systems Memory is what turns a stateless chatbot into an agent that learns. A memory system has three parts: what gets stored, how it's retrieved, and how it's forgotten or updated. Design all three deliberately. ## Overview Short-term memory is the working context: the conversation, recent tool outputs, the current plan. It's fast, exact, and bounded by the context window. Long-term memory persists across sessions: facts about the user, project state, learned preferences, past decisions. The bridge between them is retrieval — deciding what to recall and when — and write policies — deciding what deserves to be remembered. The best systems treat memory as a living document with explicit update and expiry rules, not an append-only log. ## When to use - The agent must remember user preferences or project facts between sessions. - Long conversations lose track of earlier decisions or constraints. - The agent repeats questions the user already answered. - You want the agent to learn from corrections ("I prefer X, stop doing Y"). - Multi-step work needs durable state: plans, progress, intermediate results. ## Core concepts - **Working memory**: the active context — system prompt, recent messages, scratchpad. Managed by window size, summarization, and attention to what matters now. - **Episodic memory**: records of past interactions — "on Tuesday the user asked for X and we decided Y." Useful for continuity and personalization. - **Semantic memory**: d