A few weeks ago I discovered the LLM Wiki concept proposed by Andrej Karpathy, and it immediately resonated with many of the challenges I see in complex enterprise environments.

Most organizations still struggle with fragmented knowledge:

• Software scattered across multiple repositories
• Outdated documents and wikis
• Analyst notes and undocumented processes
• Operational knowledge locked inside teams

Traditional RAG systems improve retrieval through vector similarity, but the understanding itself does not persist or evolve. The model has no memory of previous interactions and must re-process context on every new query.

LLM Wiki introduces a very different model.

     Existing Enterprise Knowledge
    --------------------------------
    - Software programs
    - Documents
    - Business concepts
    - System flows
    - Analyst insights
    - Operational knowledge

                    |
                    v

          +-------------------+
          |      LLM Wiki     |
          |-------------------|
          | Persistent        |
          | interconnected    |
          | knowledge base    |
          +-------------------+
                    ^
                    |
           new knowledge
          discovered by
             the LLM

                    |
                    v

          +-------------------+
          |        LLM        |
          |-------------------|
          | Uses wiki         |
          | articles as       |
          | contextual memory |
          +-------------------+

Instead of treating documents as static files, the LLM incrementally builds and maintains a persistent, interconnected knowledge base — essentially a continuously evolving internal Wikipedia for the organization.

The wiki doesn’t just store information. It actively:

• Connects related concepts
• Discovers hidden relationships
• Accumulates context over time
• Improves as new information is added

One of the most interesting aspects is the feedback loop:
the LLM doesn’t only consume knowledge — it can also generate new structured insights that are validated and fed back into the wiki itself.

For complex enterprise environments, especially those involving legacy systems, this approach could significantly improve:

• Reverse engineering
• Onboarding
• Impact analysis
• Documentation quality
• Long-term knowledge retention
• Legacy modernization initiatives

This feels less like “AI-powered search”
and more like building a persistent semantic memory layer for the enterprise.

Original concept:
LLM Wiki gist by Andrej Karpathy