A chatbot stores every conversation in a database and injects prior exchanges into context. A developer claims this makes it an 'agent with memory.' What is wrong with this claim, and what would need to change to make it an agent?
Memory gives a chatbot continuity, not agency. An agent must call tools, change external state, observe the result, and loop on it. A memory database is not a tool-using control loop.
Imagine a friend with a perfect diary. They remember every conversation you have ever had, so they can pick up exactly where you left off. That is impressive, but they still only talk. They cannot mail a letter for you, book a table, or check whether the train is late. They can only tell you to do those things yourself. Now imagine a personal assistant. The assistant also remembers everything, but they can actually go out, make the phone call, read the reply, and adjust the plan based on what they hear. The diary friend has memory. The assistant has agency. A chatbot with a database is the diary friend. It recalls the past, but it never reaches into the world, never sees what its actions caused, and never loops on that result to decide the next move.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
The developer's claim rests on a single confusion: the belief that memory is the defining property of an agent. It is not. Memory is about what a system retains across interactions. Agency is about whether a system can act on the world and adapt to the consequences of its actions. These are independent axes. You can have rich memory with zero agency, which is exactly the memory chatbot, and you can have agency with almost no memory, which is a stateless agent that solves a task in a few tool calls and forgets it.
The reason the confusion is so common is that memory and agency often ship together in the same product, so people assume one implies the other. A capable assistant usually remembers context and takes actions, so the two get bundled in the popular mental model. But bundling in a product is not the same as logical dependence. Pull the two apart and you can build either one without the other, which is precisely what this question is testing.
This matters in interviews because it forces a candidate to articulate the actual boundary of the agent definition rather than reciting buzzwords. The clean answer hinges on one structure: a closed loop that acts on an external environment and observes the result. Everything else, including memory, is supporting machinery that makes the loop more effective without being what makes the system an agent.
Why memory is necessary but not sufficient
Persistent memory is genuinely useful. By storing every prior exchange and injecting the relevant parts into context, the chatbot stays coherent across days or weeks. It can refer back to a decision you made last Tuesday and avoid contradicting itself. That capability solves the long-term coherence problem, which is real and hard, and most production assistants invest heavily in it through summarisation, vector stores, and retrieval over past turns.
But coherence is not action. Recalling that you wanted a report emailed does not email the report. The memory layer changes only one thing in the universe: the contents of the conversation store. Everything else, your files, your inbox, the external APIs, is untouched. The system has perfect recall and zero reach.
The sufficiency gap is the whole point. An agent is defined by a capability the chatbot lacks no matter how much it remembers. More context and longer windows scale the memory axis. They never cross over onto the agency axis. A million-token context window holding ten years of conversation is still just a very well-informed talker. It can describe the action it would take in exquisite detail, but describing an action and performing it are different operations, and only one of them is agency.
It is worth noting the reverse case too. A stateless agent with no long-term memory can still be fully agentic within a single run. It observes, calls a tool, sees the result, and loops to a goal, then discards everything. It remembers nothing across sessions yet acts on the world every turn. That asymmetry is the cleanest proof that the two properties are independent.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- ChatGPT's memory feature recalls user facts across sessions but is still a chatbot. The same product becomes agentic only in its tool-using modes like code interpreter or browsing.
- A Retrieval-Augmented Generation support bot with a full conversation store still cannot act. It retrieves and answers, but escalating a ticket requires a human or an explicit tool call to a system like Zendesk.
What an interviewer would ask next. Try answering before peeking at the approach.
QIs a system that injects retrieved documents into context, RAG, an agent? Why or why not?
Distinguish passive retrieval as context from retrieval as a tool call inside a loop. Plain RAG is one-shot and changes no external state, so it is not agentic. Agentic RAG issues retrieval as a tool action and conditions the next step on what came back.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating memory as the defining property of an agent. Memory supports coherence across turns, but agency comes from a tool-using loop that changes external state and observes the result.
60 second bullets to scan on the way to the call.
State why memory supports coherence but is not the source of agency.
Name tool execution and external state change as the missing capability.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.