5. Memory: scratchpad, summary, and long-term store

Agents need memory beyond one context window:

| Tier | Holds | Lifetime | Backed by | | --- | --- | --- | --- | | Scratchpad | current message list | one run | in-process array | | Episodic summary | compressed history of the run | run or session | model-generated summary | | Long-term | durable facts and preferences | forever | database + vector search |

Compress the scratchpad when it approaches a token threshold instead of truncating blindly:

ts
async function compress(messages: Msg[], limit = 12_000) {
  if (countTokens(messages) < limit) return messages;
  const [system, ...rest] = messages;
  const head = rest.slice(0, -6);                    // keep the last 6 turns verbatim
  const summary = await chatOnce(
    "Summarise the facts, decisions and open questions below in under 200 words.",
    head,
  );
  return [system, { role: "assistant" as const, content: "Context so far: " + summary }, ...rest.slice(-6)];
}

Long-term memory should store facts, not transcripts: `user prefers metric units` retrieves far better than a 40-turn conversation. Write facts through an explicit `remember` tool so the agent's memory is auditable and deletable — a requirement for GDPR-style deletion requests.

Exercise: add `remember(fact)` and `recall(query)` tools backed by a table with a vector column, then verify a preference stated in session one is applied in session two.

Read in another language

How to study this topic

Scoring

Check your understanding

Badges