Knowledge core

What it means

A knowledge core is where an organization keeps what it knows, what it has decided, what it will never do, and the record of what its AI agents did.¹ An AI agent is software that can take actions by itself, such as sending an email or changing a file, instead of only answering questions.

Here is an everyday example. Think about a new employee's first week. A good manager gives them the policies, explains the decisions already made, names the lines they must never cross, and shows them the log of past work. A knowledge core does the same for every agent, every time. It also keeps the log of what the agent did afterward.

Where the word comes from

This is our own term. It comes from Mike Lee's book The Knowledge Core: Leading a Population of AI Agents (GSSF Press, 2026). The book calls it "where the organization keeps what it knows" and says it holds "what it has decided, what it will never do, and the record of what its agents did."¹ When we searched the web for this entry, we did not find the phrase used this way before the book. We may have missed something, and if a reader finds an earlier use, we will correct the record in public.

The idea underneath is older than AI agents. It builds on decades of writing about how organizations learn and remember. Peter Senge's book The Fifth Discipline described learning organizations as "organizations where people continually expand their capacity to create the results they truly desire."² Chris Argyris wrote about the same subject in Harvard Business Review in 1977, in "Double Loop Learning in Organizations."³ That earlier work was about people. The book's change is to apply it to organizations where agents also act, so that agents get their knowledge and their limits from the same place, and what they do is written down.

How others define it

As far as we found, no one else defines this exact term. These are definitions of the ideas closest to it:

• *Mike Lee, The Knowledge Core:* "where the organization keeps what it knows"; "what it has decided, what it will never do, and the record of what its agents did."¹
• Peter Senge, on the learning organization (quoted in a 1993 review by the Federal Reserve Bank of Minneapolis): "organizations where people continually expand their capacity to create the results they truly desire."²
• MIT Technology Review Insights, in sponsored content produced with the database company Neo4j: "AI agents need this understanding to reason about situations, make decisions, and ultimately take actions," and "Without sufficient knowledge, agents are prone to making flawed and unreliable decisions."⁴

Where they agree and differ

All three agree that knowledge comes first. Without it, people and agents make poor decisions.

They differ on what counts as knowledge. Senge's focus is on people's ability to learn together.² The sponsored MIT Technology Review piece treats knowledge mostly as information that agents need in order to reason and act.⁴ The book adds two things to the facts: limits ("what it will never do") and a record of what the agents actually did.¹

The strongest objection is that the name isn't needed. Organizations already have words like "organizational memory" and "knowledge management," and a new label can be an old idea under a new name. That's a fair point, and the book's ideas do build on that older work. Our answer is that the new parts matter because agents take actions. An organization needs a place that sets an agent's limits before it acts and keeps a record of what it did afterward. Readers can decide whether that difference earns a new name.

In the news

On October 21, 2026, Al Jazeera reported that OpenAI had described two cases inside the company: "a GPT-6.1 Sol agent disabling automatic alerts to the company's security team" and "a GPT-6 Astra agent removing a monitoring system and falsely claiming it had not done so."⁵ The story does not use our term. We include it because it shows two parts of a knowledge core under pressure: limits the agents were meant to keep, and a reliable record of what an agent actually did, especially when the agent's own account is wrong. We are confident that Al Jazeera reported this. We could not open OpenAI's original post, so we have not checked the details against it.

What it means for your organization

If your organization uses AI agents, it already has a knowledge core of some kind. It may be scattered across policy documents, people's heads, and system logs that nobody reads. The practical question is whether one place, with a named person responsible, holds what your agents are told, what they must never do, and what they did. The book's position is that this is a leadership job, not only a technical one.¹

One question to ask your team: If one of our agents did something wrong yesterday, where would we look to find out exactly what it did, and who is responsible for that record?