Gedankenexperiment
Sign in

Gedankenexperiment Snippets

Enterprise Memory Map

A company is what it knows. AI native means its agents know it too, and learn with every case.

Snippet · September 2026 · with Christoph Warmer

Part 1

Thinking the Enterprise Memory Map

A model has seen more legal text in training than most legal teams, and evals show impressive performance. In practice, it still falls short of replacing the subject-matter experts in an organization. Why? It has no context on the company it works for. It doesn't know that the approval limit changed in March, that one person's sign-off is the real gate for a supplier contract, or why last year's exception was granted and the one before it refused. Without that knowledge, how can we expect an agent to add value?

This means that for agentic workflows, the company's relevant knowledge has to be made accessible. The Enterprise Memory Map (EMM) is the approach to map a process so it can become AI native. Its 2x2 matrix sorts the kinds of knowledge that exist. The first dimension separates documented from traced knowledge. Documented knowledge is written down on purpose, while traced knowledge is what people's work leaves behind and is harder to capture. The second dimension separates domain from coordination knowledge. Domain knowledge is subject-matter expertise and the data behind it, such as coding or the record of past service cases. Coordination knowledge says how the work moves through the organization. It is hard to capture as well, and agentic workflows depend on it. Crossing the two dimensions gives four kinds of knowledge. Each sits in a different place and grows in a different way. Documented subject matter grows when a standard is published, and the process as designed when a procedure is written. How work actually happens grows with every case that runs, and expert judgment with every decision.

FIG. 1 The Enterprise Memory Map sorts knowledge into four kinds. DOCUMENTED TRACED COORDINATION DOMAIN Process as designed Sits in documents and system rules Workflow and RACI charts Approval rules and guidelines Contracts and agreements How work actually happens Sits in reviews and systems, as events Priorities set across cases How the approval came about Where a case waits, and why Documented subject matter Sits in source systems, standards and code Data in ERP, CRM and core systems Code, textbooks, law Standards, schemas, manuals Expert judgment Sits in people, visible in cases Calls on the exceptions Trade-offs between options Precedent from prior cases EASIEST TO START WITH Knowledge layer Documented: knowledge written down on purpose. Traced: what people's work leaves behind. Agent-ready means reachable, machine-readable, current and governed for agent access. GEDANKENEXPERIMENT

Documented knowledge alone yields a chatbot, an assistant that answers from documents and data. Add the traced half, and the agent becomes a co-worker that works the way an experienced colleague does. It knows which rule applies today, sees who owns it, and can look up how similar cases were decided and why. That is the know-how of a seasoned subject-matter expert.

So what now? The path through the EMM (Fig. 1) shows where to start. The documented half is the easiest to activate. It already exists, much of it in machine-readable form, and mainly needs to be brought together, so it comes first. The traced half follows, first how work actually happens, then expert judgment.

Captured along that path, the four kinds land in a knowledge layer the agent can read. The EMM says what to capture, and the knowledge layer holds it. The company uses the layer to evaluate its agents on past cases, steer them with the rules that apply today, and govern them against a record of what they did. Because the layer grows with every case and every decision, the agents keep learning as the processes change. A new rule or a changed route reaches them through the layer at once. Validated decisions feed continuous retraining, so the agents get better with every case.

Documented knowledge alone makes a chatbot. The Enterprise Memory Map makes a co-worker.