“If That Person Leaves, It’s Over” — Are You Still Ignoring That Fear?

A manufacturing company with 30 employees. Only Mr. Yamada knows how to create estimates. Only Ms. Sato has the knack for handling complaints. Only Mr. Tanaka can close the accounting process.

This situation is familiar to many small businesses. And everyone is vaguely aware of it. “If that person leaves, this company won’t function anymore.”

Let’s create a manual. Let’s organize handover documents. We’ve been told this many times. But the reality is, there’s no time for that in a busy workplace. As a result, personalization is left unchecked, and risks quietly escalate.

Now, a technology that could fundamentally change the situation has entered the practical stage. Agent Memory — a system where AI “remembers” the context of operations and the history of decisions, storing and retrieving them when needed.

In short, AI will automatically remember the knowledge that only Mr. Yamada had in his head, making it accessible to anyone. This is how we break down personalization.

What is Agent Memory? — The Definitive Difference from “Disposable AI”

If you’ve ever used ChatGPT, you’ll understand. Once you close the chat, the AI forgets everything. Every time, you have to explain from scratch. This was the limitation of conventional AI.

Agent Memory solves this “forgetting problem”. It is a system where AI can continuously save, update, and search past interactions, operational contexts, and the rationale behind decisions.

Recent research has highlighted the concept of six-layer memory architecture. This design hierarchically categorizes memory by purpose.

Do you notice something? This is precisely the structure of a veteran employee’s mind.

Veterans are strong not just because they have knowledge. They are strong because they can make contextual judgments like, “This was the case during that project,” “This pattern works for this customer,” or “This information is old and needs verification.” Agent Memory replicates this structure in AI.

Effects in Numbers — 55% Reduction in Tokens and 20% Increase in Success Rate

Abstract discussions are meaningless without numbers. Let’s look at the data.

In one research case, a system called MAGE (Memory as Agent-Guided Exploration) produced the following results:

• Task success rate: Improved by 7.8 to 20.4 points
• Token consumption: Reduced by 55.1%

The reduction in token consumption directly translates to a decrease in API usage fees. If the monthly API cost was 100,000 yen, it would drop to approximately 45,000 yen. That’s a difference of 660,000 yen annually. For small businesses, this difference is significant.

The increase in task success rate is even more fundamental. Because AI “remembers previous failures,” it avoids repeating the same mistakes. The same structure that allows human veterans to learn from experience begins to operate within AI.

So, How Much Does It Cost? — Real Cost Estimates for Small Businesses

“I get that it’s impressive. But how much does it cost?” — This is the most important question.

To be honest, there’s no need to build a six-layer memory from scratch. The realistic option for small businesses to use right now is a combination of existing tools.

Practical 3 Steps and Cost Estimates

Step 1: Test with One Operation First (Months 1-3)

• Target: Choose one operation with the most severe personalization (like estimate creation or inquiry response)
• Method: Use Claude / ChatGPT’s project function + input internal knowledge
• Cost: 20,000 to 30,000 yen per month (API fees + tool usage fees)
• Workload: 2-3 hours per week from the responsible person
• Goal: Create a state where it can operate at 70% accuracy even without that person.

Step 2: Connect Internal Systems with MCP Server (Months 4-6)

• Use MCP (Model Context Protocol) to connect AI with internal databases and file servers
• This allows AI to reference the latest internal information while responding
• Cost: Initial setup 150,000 to 400,000 yen, monthly operation 30,000 to 50,000 yen
• Note: The failure patterns of the MCP server have been detailed in research. Connection errors, authentication failures, schema inconsistencies, etc. By deciding on a failure response flow in advance, you can significantly reduce costs during troubleshooting.

Step 3: Hierarchical Memory and Automatic Updates (From Month 7 Onwards)

• Introduce vector databases (like Pinecone, Qdrant) to enhance memory accumulation and search capabilities
• Add mechanisms for automatic archiving of old information and assigning reliability scores, thus implementing meta-memory
• Cost: 50,000 to 100,000 yen per month (database + API fees + operation)

Total Cost Comparison

The initial cost is less than one-third of the traditional approach. Moreover, it doesn’t end with just “creating it.” AI continuously learns and updates its memory within operations. It’s like having a manual that automatically stays up to date.

The Structure That Small Businesses Can “Win” With

Here’s an important point.

When large companies try to implement Agent Memory, it can take six months to a year due to security reviews, internal adjustments, and vendor selection. Just coordinating data between departments can become a major project.

Small businesses are different. If the CEO says, “Let’s do it,” they can start next week. It’s not uncommon for all company data to be in one server. There are no departmental barriers.

In other words, Agent Memory is a rare technology that works to the advantage of “smallness.” Data integration is easier, decision-making is faster, and the hurdles for company-wide deployment are lower. While large companies are busy with internal adjustments, small businesses can deploy it in three months.

Risks to Be Aware Of — “AI Memory” Is Not Omnipotent

It would be irresponsible to only hype the benefits, so let’s clarify the risks as well.

1. Hallucination Risk: AI may provide plausible answers for things it “does not remember.” The countermeasure is to manage the “confidence level of this information” in the meta-memory layer. Design it so that low-confidence responses always require human verification.

1. Information Leakage Risk: Since internal confidential information will be memorized by AI, data storage locations and access controls are essential. Decide in advance the range of information to be sent to the cloud API.

1. Dependency Risk: Relying too much on AI can lead to humans losing their judgment. Establish rules that use AI responses as “reference information,” with final decisions made by humans.

1. MCP Server Failures: Design fallback mechanisms for when the connected server goes down. At a minimum, ensure that AI can provide basic responses offline by having a local cache.

Conclusion — Start with “One Operation First”

You no longer need to spend 1 million yen on consulting to eliminate personalization.

Today, there are just three things you should do.

1. Identify the one operation with the most severe personalization in your company
2. Conduct a 30-minute interview with the veteran handling that operation and document the decision-making criteria
3. Input that into the AI project function and let a newcomer use it

With just this, you will start to see a state where it can operate at “70% accuracy.” You can then consider connecting to the MCP server and introducing a vector database to fill in the remaining 30%.

Don’t get the order wrong. Technology selection comes second. “Which personalization to break down” comes first.

The technology of Agent Memory is still evolving. But it has already reached a “usable level.” To avoid looking back in six months and saying, “I should have started then,” I urge you to decide on your first operation this week.