1️⃣ STATE — The Living Memory

What is State?

State is the single source of truth in LangGraph.

It is the data object that flows through every node.

Mental Model

🧠 State = Agent’s brain

Everything the agent knows or has produced so far lives here.

What State Contains

State can contain:

• Messages (conversation history)

• Tool outputs

• Flags

• Counters

• Intermediate results

• Retrieved documents

Example (Conceptual)

state = {
  "messages": [...],
  "documents": [...],
  "step_count": 3,
  "is_finished": False
}

📌 State is mutable across nodes
📌 State grows as the graph runs

2️⃣ TypedDict — The Blueprint of State

What is TypedDict?

TypedDict defines WHAT keys the state can have and their types.

It does not store data.
It describes data.

Mental Model

📐 TypedDict = Architectural blueprint

It tells LangGraph:

• Which fields exist

• What type each field must be

Why TypedDict Exists

• Prevents accidental state corruption

• Makes large graphs predictable

• Enables static checks & IDE support

• Forces discipline in agent design

Example (Conceptual)

class AgentState(TypedDict):
    messages: list
    documents: list
    is_finished: bool

📌 TypedDict never changes at runtime
📌 It enforces structure, not behavior

3️⃣ Messages — Reasoning Units

What are Messages?

Messages are atomic units of conversation.

Each message is an instance of:

• HumanMessage

• AIMessage

• SystemMessage

• ToolMessage

All inherit from BaseMessage.

Mental Model

💬 Messages = Thoughts and memory fragments

Agents think by appending messages.

What Messages Do

Messages:

• Carry user intent

• Store LLM reasoning

• Preserve tool results

• Drive routing decisions

• Enable replay & explainability

Example (Conceptual)

messages = [
  HumanMessage("Explain RAG"),
  AIMessage("Let me explain..."),
  ToolMessage("Search results")
]

📌 Messages are append-only
📌 They form the reasoning timeline

🔁 How They Work Together (Very Important)

Flow Relationship

TypedDict → defines State structure
State     → holds Messages + data
Messages  → drive reasoning & control

Runtime Reality

• TypedDict stays constant

• State evolves

• Messages accumulate

⚙️ Who Does What (Quick Comparison)

🧠 Agentic RAG Perspective

• TypedDict ensures RAG fields exist

• State carries retrieved chunks

• Messages store:

  • user query

  • retrieval results

  • reasoning steps

  • final answer

📌 Without messages, RAG is just search.
📌 Without state, agents forget.
📌 Without TypedDict, systems break.

🥇 Interview Gold Line

“TypedDict defines the shape of state, state carries the data, and messages are the reasoning units that drive LangGraph execution.”

🧠 One-Line Summary

TypedDict is the contract, State is the memory, Messages are the thoughts.