Giving Your Agent Long-Term Memory

An agent with no memory helps once. An agent with the right memory gets better every week.

The three layers, and when to add each

• Working memory — the current conversation. You already have it; keep it lean.
• Semantic memory — durable facts ("this user prefers concise replies"). Add this first; a single vector store covers most needs.
• Episodic memory — specific past events ("last Tuesday they asked about refunds"). Add only when a real task needs recall of *when*.

A minimal setup that works

1. Summarize each session into a few durable facts.
2. Store those facts with embeddings.
3. On each new task, retrieve the most relevant facts and prepend them.

The traps

• Storing everything — memory becomes noise; store distilled facts, not transcripts.
• Never forgetting — let stale or contradicted facts expire.
• No retrieval test — if you can't see what was recalled, you can't debug it.

Start with semantic-only. Most teams never need more.