การจัดการหน่วยความจำและสถานะสำหรับเอเจนต์

LLM agents are designed to perform complex tasks. But by default, LLMs are stateless! This means they don't "remember" past interactions.

For an agent to have a meaningful conversation or complete multi-step tasks, it needs a way to recall previous information. This is where memory comes in.

Agent memory can be categorized into two main types:

• Short-Term Memory: This is like a human's working memory. It holds recent, relevant information for immediate use, often the raw conversation history.
• Long-Term Memory: Stores information over extended periods, like past conversations or learned facts. It helps agents retain knowledge beyond the current interaction.

The simplest form of short-term memory for an agent is a conversational buffer. It stores the raw exchange of messages between the user and the agent.

When generating a new response, the agent is prompted with the entire buffer, allowing it to maintain context for the current turn.

Here's a simple Python example simulating a conversational buffer. Each new message is added to a list that represents the agent's memory.

Raw conversation buffers grow quickly, hitting LLM context window limits and increasing costs. Summarization is a key technique for long-term memory.

Instead of sending the full history, we periodically summarize older parts of the conversation, keeping the key points while discarding verbose details.

This example shows how a summary might be generated and used. In a real application, an LLM would create the summary, but here we simulate it.

Beyond just remembering conversation, agents need to manage state to complete multi-step tasks. This means tracking progress, decisions made, and information collected.

Think of it like a checklist or a finite state machine for the agent's goal. For example, booking a flight requires knowing source, destination, and dates.

Consider an agent helping book a flight. Its internal state might include:

• Status: "collecting_info", "searching_flights", "booking_confirmed"
• Departure City: "London"
• Destination City: "Paris"
• Departure Date: "2024-12-25"
• Number of Passengers: "1"

The agent updates this state as it gathers information from the user.

Managing agent memory and state introduces several key challenges:

• Context Window Limits: How much information can the LLM process at once?
• Cost: Longer contexts mean more tokens and higher API costs.
• Retrieval: For very long-term memory (e.g., a knowledge base), how does the agent efficiently find the most relevant past information?
• Consistency: Ensuring the agent's state accurately reflects the task's progress and collected data.

Let's check your understanding of agent memory and state management techniques.

In this lesson, we explored how memory and state are essential for LLM agents to engage in long-term conversations and complete complex tasks.

• We covered short-term conversational buffers and long-term summarization techniques.
• We also looked at how state management helps agents track progress through multi-step processes.

These techniques are fundamental for building truly capable and persistent AI agents.