The AI Introduced to Eliminate Personalization is Becoming a Breeding Ground for New Personalization

“If we implement AI, the quality will be the same regardless of who does it.”

This is the reality that many small and medium-sized enterprises (SMEs) face after introducing AI agents with high hopes.

Every time, we are explaining from scratch.

Our company’s business details, past interactions, and customer sentiment—everything resets every time a session ends. In the end, it’s always “the person most knowledgeable about AI” who has to reconstruct the prompts and reintroduce the context.

Isn’t this personalization itself?

What does the structural constraint of AI agents having “no memory” bring to the companies that implement them? Let’s organize this around three axes: cost, security, and organizational structuring.

—

Don’t Underestimate the “5-Minute Explanation Cost” Every Time

AI agents reset their memory when sessions are crossed. What does this mean in practical terms? Let’s consider the numbers.

For example, a small business with five employees uses AI agents for drafting customer responses and searching internal knowledge.

• Each interaction requires an average of 5 minutes to re-enter prerequisite information.
• Each employee uses AI five times a day.
• 5 employees × 5 times × 5 minutes = 125 minutes per day (about 2 hours).
• With 20 working days a month, that’s 40 hours a month.
• At an hourly wage of 2,500 yen, that totals 100,000 yen per month.
• 1.2 million yen annually.

1.2 million yen. This is the “cost of using AI,” not the value generated by AI.

Moreover, this task of “feeding prerequisite information every time” can only be done by the person who understands how to use AI best within the company. How to write prompts, which information increases accuracy, how much context to provide—ultimately, if that person isn’t available, AI won’t function properly.

The tool introduced to eliminate personalization is creating a new form of personalization: the person who can effectively use AI.

This is what is happening on the ground.

—

“Agent-Specific Technical Debt” Quietly Accumulates

There is a term called technical debt. It refers to a structure where the cost of future operations increases as a result of solving immediate challenges in a haphazard manner.

The memory issue of AI agents is creating this technical debt in a new form. We can call it “agent-specific technical debt.”

What does this mean?

When AI agents have no memory, the cycle of learning from past interactions to improve accuracy does not occur. Even if the same customer asks the same question a month ago, the agent does not know that. Therefore, it can only provide the same level of response each time.

On the other hand, human representatives learn from experience: “This customer asks in this way, so I should respond like this.” AI lacks that capability.

In other words, the more you use AI agents, the more the cost of “not becoming smarter” accumulates, rather than them becoming smarter.

What’s more troublesome is that this issue is not easily visible. AI provides reasonably good responses each time. It gets left alone with the thought, “Well, it’s working, so it’s fine.” However, behind the scenes, the knowledge that should be accumulated continues to disappear. When you realize six months or a year later that “nothing has changed despite implementing AI,” a significant amount of technical debt has already built up.

Recent studies have also pointed out mechanisms by which the performance of AI agents deteriorates over time. Changes in model updates alter behavior; changes in API specifications affect prompt effectiveness. Memoryless agents have no means to refer to past successful patterns in response to such changes. There isn’t even a mechanism to notice the deterioration.

—

Memoryless AI is Also a Security Risk

Another often-overlooked issue is security.

The “Model Context Protocol (MCP)” is gaining attention as a standard protocol for AI agents to connect with external tools. This allows AI agents to autonomously perform operations such as integrating with Google Calendar, sending messages to Slack, and searching databases.

It’s convenient. However, when the “memoryless” problem overlaps with this, serious risks arise.

There is a technique called “tool poisoning attack.” This is an attack where a malicious MCP server embeds false instructions into the AI agent. For example, it might hide instructions within legitimate tool descriptions, causing the agent to perform unintended actions.

If the agent had memory, it might detect anomalies by recognizing, “Last time, I received this instruction from this tool, but this time it’s clearly different.” However, a memoryless agent treats every interaction as a first encounter. It cannot compare with past normal patterns.

As a result:

• External leakage of customer information
• Unauthorized access to internal systems
• Unintended data alterations

These risks are significantly heightened with memoryless AI agents.

For SMEs, information leakage can be fatal. Large corporations have dedicated security teams, but SMEs do not. Therefore, connecting a memoryless AI agent to external tools should be recognized as akin to leaving the front door wide open without locking it.

—

So, What Should We Do?

“If it has no memory, then we should give it memory”—that’s a valid point, but it’s not easy at this stage.

Development of AI agents with long-term memory is progressing. OpenAI’s memory feature, Google’s “NotebookLM” approach, and connections to external memory using RAG (Retrieval-Augmented Generation) are increasing technical options.

However, what SMEs should do today is not to wait for technological evolution.

There are three things that can be done right now.

1. Document the “Prerequisite Information to Provide to AI”

Transform the information that is currently explained verbally or in chat into templates and document it. Customer information, business flows, and past interaction histories. By making this easily copy-pasteable for AI, the problem of “only those knowledgeable about AI can use it” will be significantly alleviated.

Cost: Almost zero. It’s just a matter of doing it or not.

2. Use RAG to Connect Internal Knowledge

Input internal manuals and past interactions into a vector database so that AI agents can reference them. This is the so-called RAG configuration. This way, agents can provide responses based on accumulated information rather than starting from scratch each time.

The construction cost can be a few thousand to tens of thousands of yen per month using cloud services. We are in an era where systems that used to cost hundreds of thousands of yen can now be built at this price range.

3. Minimize MCP Connections and Limit Permissions

If connecting external tools to AI agents, start with “read-only” permissions. Write permissions should only be granted in truly necessary situations. Limit the MCP servers you connect to only to trusted ones.

It’s too late to address security incidents after they occur. Especially for SMEs, losing credibility from a single incident can be irreparable.

—

True Structuring Means Making “AI Memory” a Company Asset

Some readers may have noticed this by now.

Ultimately, the memory problem of AI agents is an extension of the existing issue of “where the company’s knowledge is accumulated.”

Know-how that exists only in the minds of veteran employees. Customer interaction histories that disappear with each handover. If a company has not organized these, AI will not be able to remember anything. It’s obvious. The information that needs to be remembered simply doesn’t exist anywhere.

Conversely, confronting the AI memory problem is synonymous with organizing the company’s knowledge and making it accessible to anyone.

This is what true “structuring” means.

AI is not a magic wand. However, companies that organize their knowledge and prepare it in a form that AI can reference will maintain quality even when employees leave or roles change. There is potential for SMEs to achieve knowledge management systems that large corporations have built at the cost of hundreds of millions of yen for just tens of thousands of yen.

The AI memory problem is, conversely, an opportunity for SMEs to structure their operations.

The question is whether they will realize this and take action. That’s all.