Can Your Business Operate if Your Veteran Employees Leave?

If you are a small or medium-sized enterprise owner, you have likely experienced this fear at least once.

Only person A knows the key points for estimates. Only person B can determine how to handle complaints. Only person C remembers the negotiation patterns with this client. — Such “know-how that exists only in their heads” holds the lifeline of the company.

When trying to create a manual, the judgments of veteran employees are often tacit knowledge. There are too many instances of “I just do this when this happens” that cannot be documented. As a result, personalization is left unchecked.

This structure has now entered a phase where it can be technically dismantled. Moreover, for less than 50,000 yen a month.

What Happened: Tencent Open-Sourced “AI Memory”

Tencent’s “TencentDB Agent Memory v2.0,” which will be open-sourced in 2025, is a team-level memory hub for AI agents. It is licensed under the MIT License and can be used for free.

In a nutshell, it is a system that “continues to remember all business interactions—conversations, documents, and code—in a form that AI can reuse.”

Specifically, it automatically converts accumulated information into four memory assets:

• Chat Memory: Retains context from past conversation histories, eliminating the need for repetitive explanations.
• Skills: Extracts and makes successful judgment patterns and procedures reusable.
• LLM-Wiki: Structures internal knowledge for AI to reference.
• Code Graph: Remembers code dependencies and change histories to streamline development tasks.

The key point is that “once something has been explained, it does not need to be explained again.” Even when starting a new session, the context of past projects is carried over. In human terms, it’s like having a “new employee who can contribute from day one after reading the handover materials.”

Implementation is as simple as one command with Docker. It can be hosted on your own server, eliminating the need to expose confidential information externally. With just one engineer, an operational environment can be set up in half a day.

The Essence is a Change in Cost Structure

What we want to consider here is not merely the emergence of a “convenient tool.” It is a structural change where the cost of knowledge management has dramatically decreased.

What has been required to eliminate personalization until now?

• Interviews with veteran employees: dozens of hours.
• Outsourcing manual creation: 500,000 to 3,000,000 yen.
• Implementing knowledge management systems: 1,000,000 to 5,000,000 yen annually.
• Even then, “tacit knowledge” cannot be fully documented.

Even after spending hundreds of thousands of yen, manuals often go unread and systems become mere shells. For small and medium-sized enterprises, this has been an area they wanted to tackle but couldn’t.

Now, this is how it changes:

• TencentDB Agent Memory: Free (OSS)
• Infrastructure costs for self-hosting or cloud: 10,000 to 30,000 yen per month.
• API usage fees for LLMs (like GPT-4o or Claude): 10,000 to 30,000 yen per month.
• Initial setup labor: 1 engineer for 1 to 2 days.

Less than 50,000 yen per month. Initial costs are almost zero.

300,000 yen has become 50,000 yen. This difference is not about “whether to do it or not,” but rather about “whether you know about it or not.”

There’s More Ahead: Research on AI That “Learns and Evolves on Its Own”

While Tencent’s memory hub is about systematizing “memory,” there is another noteworthy movement. Research on AI agents that acquire and evolve skills through practical tasks.

In the framework called “EvoHarness-RL,” agents learn basic judgment patterns offline and then improve their abilities in real-time by updating external states while executing actual tasks. It’s like an “AI that grows through on-the-job training.”

What happens when this becomes practical?

For example, in manufacturing estimation tasks. Initially, past estimation data and judgment criteria from veteran employees are accumulated in the memory hub. The AI creates estimates based on that reference. Up to this point, it’s about “reusing memory.”

However, a self-evolving agent can incorporate feedback from actual orders won and lost, autonomously improving the accuracy of estimates. It could even surpass the intuition of veteran employees.

The important point is that this evolution is shared across the entire team. Skills learned by one agent can be disseminated to other agents through the memory hub. This is the exact opposite of personalization — a state where “anyone can make the best move” is becoming technically feasible.

Why Small and Medium-Sized Enterprises Can Ride This Wave

“But we don’t have engineers” or “This is for large companies” — you might think. On the contrary.

When large companies try to implement such tools, they face security reviews, approvals from IT departments, vendor selection, PoCs, and internal approvals… It takes at least six months.

In a small or medium-sized enterprise, if the president says, “Let’s do it,” they can start moving by next week. They can set up a Docker environment and first run the memory hub for one task — for example, handling inquiries or creating estimates. They can test it for a month and see the effects. If it doesn’t work, they can stop. At 50,000 yen a month, the withdrawal cost is almost zero.

This “speed of decision-making” and “ability to start small” is the greatest weapon of small and medium-sized enterprises. While large companies are going through internal approvals, they can conduct three rounds of experiments.

Specifically, Where to Start

“This sounds interesting, but what should I actually do?” — It’s meaningless unless you delve into this.

Step 1: Choose one task that is personalized
Don’t try to do everything. Identify just one task that would stop if this person were absent. It could be estimates, complaint handling, order decisions, anything.

Step 2: Start accumulating “judgment logs” for that task
Record the conversations, emails, and chats of veteran employees during their work (with their consent). There’s no need to create a perfect manual. Raw interactions are sufficient.

Step 3: Set up and feed TencentDB Agent Memory
Launch it with one Docker command. Input the accumulated logs into the memory hub. The AI will automatically structure them into skills and knowledge.

Step 4: Try using it in practice
Have new employees or others perform that task through the AI agent. Monitor the accuracy and provide feedback.

Step 5: Measure the effects
After a month, compare response times, error rates, and the number of questions directed to veteran employees. Make decisions based on numbers.

Total investment up to this point: less than 50,000 yen + a few days of engineer labor. It’s an amount that won’t hurt even if it fails.

“Personalization” is No Longer an Excuse

To be honest, if you ask whether this technology is perfect, it’s still a work in progress. It cannot fully replicate the tacit knowledge of veteran employees. Complex judgments require human review, and there may be cases where it doesn’t fit certain tasks.

However, the era of “there are no means to eliminate personalization” is over.

With an open-source memory hub, running costs under 50,000 yen, and implementation via a single Docker command, the technical barriers have dramatically lowered. Now, it’s just a matter of “whether to do it or not.”

Can you systematize the knowledge in the minds of veteran employees before they retire? This is not a technical issue but a management decision.

Is there any reason not to conduct an experiment that can be started for 50,000 yen a month?