Ep 13: Retaining Memories — Window Buffer Memory & Conversation State Persistence

Why Agents Need Memory

LLMs are stateless — each call is brand new. Without Memory:

sequenceDiagram
    participant User as 👤 User
    participant AI as 🤖 Memoryless Agent
    User->>AI: My name is Alice
    AI-->>User: Nice to meet you, Alice!
    User->>AI: What's my name?
    AI-->>User: Sorry, I don't know your name. 😅

Memory nodes inject conversation history into the prompt before each LLM call, making the model "appear" to remember.

1. Window Buffer Memory

Maintains a fixed-size sliding window of the most recent N messages.

graph TB
    subgraph "Window Buffer (size = 6)"
        subgraph "Turn 1"
            M1["👤 Hello"] --> M2["🤖 Hi!"]
        end
        subgraph "Turn 2"
            M3["👤 Weather?"] --> M4["🤖 Sunny 28°C"]
        end
        subgraph "Turn 3"
            M5["👤 Tomorrow?"] --> M6["🤖 Cloudy"]
        end
        subgraph "Turn 4 (window slides!)"
            M7["👤 Day after?"] --> M8["🤖 Rainy"]
        end
    end
    M1 -.->|"❌ Evicted"| Trash[🗑️]
    M2 -.->|"❌ Evicted"| Trash
    style Trash fill:#ef4444,stroke:#dc2626,color:#fff

Window Size Guide

2. Session Isolation

graph TB
    CT[💬 Chat Trigger]
    CT -->|"sessionId: alice"| Agent[🤖 AI Agent]
    CT -->|"sessionId: bob"| Agent
    Agent --> Memory[💾 Memory]
    Memory --> S1["📂 alice: [hello, weather...]"]
    Memory --> S2["📂 bob: [help me code...]"]
    style Memory fill:#22c55e,stroke:#16a34a,color:#fff

3. How Memory Injects Into LLM Calls

// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
// WITHOUT Memory:
// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
const messages = [
  { role: "system", content: "You are an AI assistant" },
  { role: "user", content: "What about tomorrow?" }  // No context!
];

// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
// WITH Memory (auto-injected):
// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
const messages = [
  { role: "system", content: "You are an AI assistant" },
  { role: "user", content: "Weather in Beijing?" },     // History
  { role: "assistant", content: "Sunny, 28°C" },        // History
  { role: "user", content: "What about tomorrow?" }      // Current
];
// Model now knows "tomorrow" means Beijing weather

Memory Type Comparison

Next Episode

In Ep 14, we equip the Agent with real external Tools — enabling it to not just "talk" but "act": calculate, search, and call APIs.