Memory

How Memory Works

Agents use a sliding-window memory to maintain conversation context. The system message is always preserved. Older messages are dropped when the window is full.

[system] Always preserved
[user] message 1     <- may be dropped if window full
[assistant] reply 1  <- may be dropped if window full
...
[user] latest        <- always kept
[assistant] latest   <- always kept

Configuration

const agent = new Agent({
  name: "Chat Bot",
  instructions: "You are helpful.",
  memory: { maxMessages: 100 },
  transport: schift.transport,
});

Option	Type	Default	Description
maxMessages	number	50	Maximum messages to retain
transformContext	(messages: ChatMessage[]) => ChatMessage[]	—	Transform messages before each LLM call

Context Transform

Use transformContext to modify messages before they’re sent to the LLM. This is useful for summarizing older messages, injecting context, or filtering:

const agent = new Agent({
  memory: {
    maxMessages: 100,
    transformContext: (messages) => {
      // Truncate long tool results to save tokens
      return messages.map((m) =>
        m.role === "tool" && m.content.length > 2000
          ? { ...m, content: m.content.slice(0, 2000) + "..." }
          : m,
      );
    },
  },
  ...
});

Stateless by Default

Each agent.run() call creates a fresh memory. The system message and user input are added automatically.

For multi-turn conversations, manage conversation state externally:

// Simple multi-turn pattern
const messages: string[] = [];

async function chat(userMessage: string) {
  messages.push(userMessage);
  const result = await agent.run(messages.join("\n---\n"));
  return result.output;
}

Memory and Tool Calls

During a ReAct loop, tool calls and their results are stored in memory so the LLM can reference previous tool outputs:

[system] instructions
[user] "What's the weather?"
[assistant] [Tool call: get_weather({"city": "Seoul"})]
[tool] {"temperature": 22}
[assistant] "It's 22C in Seoul."  <- final answer