The Knowledge Department
Announcing a research and systems project exploring how AI agents can safely create, maintain, and govern shared organizational knowledge.
I've been pulling on a thread for a while now, and it's time to talk about where it's going.
If you've been following along, you might have noticed a pattern in the things I've been building. semanticwiki grew out of a fascination with how agents can digest and represent codebases as structured knowledge. Build an Agent Workshop came from wanting a better framework for defining specialized agents with clear responsibilities. Both of these are really knowledge engineering problems at heart: how do you capture what matters, structure it meaningfully, and make it useful for both humans and machines?
The latest entry in this throughline is The Knowledge Department.
What It Is
The Knowledge Department (TKD) is a research and systems project exploring a question I think is underexplored:
How do AI systems collaborate with humans to maintain trustworthy organizational knowledge over time?
The tagline is "Institutional Memory for the Agentic Era," and that captures it well. Inside organizations, knowledge is messy. Policies change. Documents contradict each other. Ownership shifts. The things people just know slowly fade as teams turn over. This is already a hard problem for humans alone, and it gets harder when AI agents enter the picture as collaborators who both consume and contribute to that shared knowledge.
TKD treats organizational knowledge as a living system that needs to be curated, validated, versioned, and audited.
The core idea is knowledge governance for AI systems. Instead of letting agents freely read and write to shared memory, TKD routes knowledge contributions through specialized AI custodians, each with a narrow, well-defined job. An Archivist categorizes incoming knowledge. A Validator checks claims and flags contradictions. Additional roles for reconciliation and gap detection are on the roadmap. Think of it like a pull request workflow, but for organizational facts instead of code.
The Shadow Clone Problem
If you've ever watched Naruto, there's a concept that maps surprisingly well to what TKD is trying to solve.
When Naruto creates shadow clones, each clone goes off independently, trains, fights, or explores. When a clone disperses, all of its experience and knowledge flows back to the original Naruto. It's a powerful shortcut for learning, but there's a catch: if multiple clones learn contradictory things, or if one clone picks up bad habits, all of that comes back too. The original has to make sense of it all.
This is essentially the problem organizations face as they deploy AI agents at scale. You might have dozens of agents working across different teams, each picking up context, making decisions, and writing things down. When those contributions flow back into the shared knowledge base, someone (or something) needs to sort out what's trustworthy, what conflicts, and what's stale. Without that layer of governance, you get a mess of contradictory "facts" that agents then build on top of, compounding the problem.
TKD is that governance layer. It's the system that receives what the clones learned and makes sure the original body of knowledge stays coherent.
Why I'm Excited About This
This sits at the intersection of things I care about: agent architecture, knowledge representation, evaluation design, and building AI systems that are genuinely trustworthy, not just capable. TKD includes a simulated enterprise environment called Watership for rigorous evaluation, a progression from prompt-based policies to reinforcement learning, and a focus on documenting how and when agents game reward functions. That last part is especially important as these systems scale.
The project is in its early stages. I'm currently in Phase 1: architectural planning and evaluation framework design. There's a full roadmap on the plan page if you want the details.
Follow Along
I'll be writing about this as it develops. You can filter posts on this blog by the the-knowledge-department tag to follow the journey. I'll share what works, what doesn't, and what I learn along the way.
If you're building AI agents and want to help shape the evaluation framework, there's a research preview waitlist on the site.
More soon.