A Company That Sees a 20% Drop in Sales the Month After a Veteran Leaves

If you are a small business owner, you have probably thought at least once, “If that person leaves, we won’t be able to operate.”

According to a survey by the Ministry of Economy, Trade and Industry, about 70% of small businesses report that “business knowledge is concentrated in specific employees.” This is what we call personalization. As a countermeasure, many companies focus on “creating manuals.” But let’s be honest: I hardly know of any company where personalization has been resolved through manuals.

Why is that? The true value of veteran employees lies not in the “procedures” that can be documented but in the “judgments” that cannot be written down. “It’s better to say this to this customer,” or “If the color of this material is slightly different, adjust it this way”—such tacit knowledge is often something even the person themselves cannot articulate.

However, by 2025, the answer to this problem is beginning to change. The AI “memory” function is rapidly evolving, making it increasingly feasible to digitally accumulate the very judgment patterns of veteran employees.

The Real Reason Manuals Fail to Solve the Problem

First, let’s clarify why manuals as a countermeasure to personalization do not work.

First, the cost of creation is high. Cataloging and documenting the work of veteran employees can take hundreds of hours. When converted to hourly wages, this amounts to an investment of several hundred thousand to a million yen. Moreover, the veteran employee themselves is often tied up in this task, causing regular operations to stall.

Second, they quickly become outdated. Business processes change daily. Who is going to maintain a manual that becomes outdated the moment it is created? In many cases, no one does.

Third, tacit knowledge cannot be documented in the first place. “Stop the machine when you hear this sound,” or “Lower the temperature when you smell this odor.” Judgments based on the five senses and exceptions based on experience cannot be replicated just by reading, no matter how carefully they are written.

In other words, manuals only address the surface of the personalization problem. A different approach is necessary to tackle the core issue of “reproducing judgments.”

What is AI Memory?—The Shock of the “Dual Memory Framework”

Recently, a memory structure of AI known as the “Dual Memory Framework” has gained attention.

This framework divides the memory of AI agents into two layers. One is “Progress Memory,” which extracts semantic blueprints from successful work patterns to guide the execution of the next task. The other is “Feasibility Memory,” which generates verification rules from past failure patterns and checks for logical validity.

If we compare it to the human brain, Progress Memory is the “memory of successful methods,” while Feasibility Memory is the “memory of things not to do.” The judgment ability possessed by veteran employees is precisely a combination of these two types of memory.

AI agents equipped with this dual memory demonstrate significantly superior performance in long-term task execution compared to conventional AI. The “infinite loop of trial and error” and “deviation from goals” that traditional AI agents often fall into have been greatly reduced by this framework.

Having “Another Veteran” for Just a Few Thousand Yen a Month

Another practical aspect worth noting is the evolution of personal AI private memory systems.

The system called “Opal” remembers user activities over the long term and structures them as a lightweight knowledge graph. It allows for information retrieval that understands context, rather than just keyword searches. Research indicates that accuracy has improved by 14% compared to conventional methods, throughput has increased 29 times, and costs have been reduced to one-fifteenth.

The implications of these numbers are significant. Previously, constructing a company’s knowledge management system would require an initial investment of several million yen and monthly operational costs of tens of thousands of yen. With a cloud-based AI memory system, however, it is now possible to achieve this for a range of a few thousand to several tens of thousands of yen per month.

A concrete image of its application is as follows: Veteran employees engage in dialogue with an AI assistant during their daily tasks. “I remember a similar case before. We moved the deadline up by a week then,” or “This material warps when the humidity is high, so I’ll lower the setting by 0.5 degrees today.” Such dialogues accumulate in the AI’s memory.

Even after a veteran leaves, successors can ask the AI, “Have there been any past cases similar to this one?” The AI can respond based on accumulated judgment patterns, saying, “Last time, we moved the deadline up, and customer satisfaction increased.” This is the contextual wisdom that will never be written in a manual.

The Essence of Business Succession is the “Inheritance of Judgment”

According to data from the Small and Medium Enterprise Agency, about 1.27 million small businesses are expected to face a lack of successors by 2025. The issue of business succession is not merely about transferring stocks or assets. The essence lies in the “inheritance of management judgment.”

The relationships with clients built over 30 years by the previous generation, the key points in price negotiations, and the patterns of response during troubles—if these are not passed on, even if the business is formally succeeded, the actual corporate value will be diminished.

AI memory could become the first practical means to technically support this “inheritance of judgment.” Of course, AI cannot perfectly reproduce all tacit knowledge. However, there is a world of difference between “nothing remains” and “70% of judgments are left in a referenceable form.”

Is This Enough?—What You Can Start Doing Today

Some business owners might think, “AI memory is still a distant prospect.” But consider this: It has only been two and a half years since ChatGPT was introduced. In that time, AI has evolved from a “text generation tool” to a “memory-holding agent.”

Here are three things you can start doing today:

1. Start taking “judgment logs” from veteran employees. You don’t need an elaborate system. Just have them write one line in their daily reports about “a situation where they made a different judgment today.” This will become the learning data for future AI memory.

2. Incorporate conversational AI tools into your operations. Whether it’s ChatGPT, Claude, or an internal chatbot, creating a habit where veterans engage in dialogue with AI while progressing through their tasks will naturally lead to the digital recording of their judgment processes.

3. Inventory “who would be in trouble if they left.” Identify high-risk areas of personalization and prioritize them. There’s no need to AI-enable everything at once. Start with the one or two tasks that carry the highest risk.

What kills personalization is not thick manuals but AI memory that absorbs the judgment patterns of veterans daily and retrieves them in the necessary form when needed. That technology is already within reach.

The question is whether to act “now” while veterans are still in the company. Once they leave, their knowledge cannot be regained.