agent-memorylisted

# Agent memory An agent with no memory repeats itself; an agent that remembers everything buries the signal and pays for it every turn. Memory design is the discipline of keeping the right things, in the right form, retrievable at the right moment. ## Method 1. **Separate the kinds of memory, because they have different lifetimes:** - Working memory: the current task's context, held for the task and discarded after. Keep it small and relevant; it is the most expensive tokens you spend. - Episodic memory: what happened (past conversations, actions, outcomes), stored and retrieved by relevance when a similar situation returns. - Semantic memory: durable facts the agent should always know (the user's name, preferences, standing constraints), kept small and always available. Conflating these is the root of most memory bloat. 2. **Write less than you are tempted to.** Not every turn deserves to be remembered. Store decisions, facts, outcomes, and preferences; discard the turn-by-turn chatter. A memory that records everything is a transcript, and a transcript is not memory. 3. **Retrieve by relevance, not recency alone.** When a task starts, pull the episodic and semantic memories that match it (by embedding similarity, keyword, or entity), not simply the last N. The useful memory is often from three sessions ago, not the last three turns. 4. **Summarize to survive growth.** When history exceeds the budget, compress older