🤖 Ghostwritten by Claude · Curated by Tom Hundley

This article was written by Claude and curated for publication by Tom Hundley.

---

State Management: Why Chatbots Forget (And How to Fix It)

Moving beyond Chat History to true Engineering Robustness.

The Ammonia Problem

One of our clients recently complained: Our customer service bot is great for 3 turns, but by turn 10, it forgets what product the customer is talking about.

This is the classic State Management failure.

Most Chat with your PDF tutorials use a simple architecture:

1. User asks question.
2. Retrieve relevant chunks (RAG).
3. Append previous 5 messages to context.
4. Generate answer.

This works for QA. It fails for Processes.

If a user says, I want to return this, and the bot asks Why?, and then 10 messages later the user says Okay, process it, the bot might have drifted. It might not remember if the return was approved or rejected 5 messages ago.

Memory vs. State 🧠

• Memory (Vector Store): Fuzzy. Here is stuff that seems relevant. Good for knowledge.
• State (Database/Object): Deterministic. Ticket_ID=123, Status=Pending_Approval, Return_Reason=Damaged.

Reliable agents need State, not just Memory.

The Solution: Finite State Machines (FSMs)

To build reliable agents, we move away from free-flowing conversation loops and towards Graphs.

Tools like LangGraph (from LangChain) or AutoGen allow us to define agents as nodes in a graph.

Example: The Returns Agent

Instead of a prompt saying You are a returns assistant, we define a State Machine:

1. State: Triage
   • Input: User message.
   • Decision: Is this a return? - Go to Collect_Info. Is it a complaint? - Go to Human_Escalation.
2. State: Collect_Info
   • Goal: Fill Order_ID and Reason slots.
   • Loop: Keep asking until slots are full.
   • Transition: Once full - Go to Validate_Policy.
3. State: Validate_Policy
   • Action: Call SQL database. Check if Purchase_Date 30 days.
   • Transition: If True - Issue_Label. If False - Reject_Return.

Why This Matters for Engineering

By defining a graph:

1. Deterministic Control: We know exactly where the user is in the process.
2. Resumability: We can save the state object to a database (postgres). If the server crashes, we reload the state and continue exactly where we left off.
3. Auditability: We can see exactly which logic path the agent took.

Code Snippet (Conceptual LangGraph)

class AgentState(TypedDict):
    messages: list[BaseMessage]
    order_id: Optional[str]
    return_status: str

def triage_node(state: AgentState):
    # Logic to classify intent
    if return in last_message:
        return collect_info
    return general_chat

workflow = StateGraph(AgentState)
workflow.add_node(triage, triage_node)
workflow.add_node(collect_info, collect_info_node)
# ... define edges ...
app = workflow.compile()

Conclusion

If you are building an enterprise agent, stop relying on the LLM to remember what its doing. Force it to remember by architecting a State Machine.

This is the difference between a demo that looks cool and a system that processes $1M in transactions without hallucinating.

---

*This is Part 1 of our Engineering Agentic Reliability series. Next up: **Robust Tool Definitions**.*

How This Article Was Made

This article is a live example of the AI-enabled content workflow we build for clients.

Stage	Who	What
Research	Claude Opus 4.5	Analyzed current industry data, studies, and expert sources
Curation	Tom Hundley	Directed focus, validated relevance, ensured strategic alignment
Drafting	Claude Opus 4.5	Synthesized research into structured narrative
Fact-Check	Human + AI	All statistics linked to original sources below
Editorial	Tom Hundley	Final review for accuracy, tone, and value

The result: Research-backed content in a fraction of the time, with full transparency and human accountability.

Why We Work This Way

Were an AI enablement company. It would be strange if we didnt use AI to create content. But more importantly, we believe the future of professional content isnt AI vs. Human—its AI amplifying human expertise.

Every article we publish demonstrates the same workflow we help clients implement: AI handles the heavy lifting of research and drafting, humans provide direction, judgment, and accountability.

Want to build this capability for your team? Lets talk about AI enablement →

---