A small company’s collective memory is usually a graveyard of folders. You’ve got project briefs decaying in Google Drive. Brand books trapped in PDFs from a year nobody wants to admit. Invoices lurking in a shadowy subtree the new intern is absolutely not supposed to open.
Humans learn to navigate this mess through slow, tribal apprenticeship. AI agents? They just learn from whatever fragmented paragraphs someone blindly pasted into the chat box last Tuesday.
The result is the exact same onboarding failure for both humans and machines: not enough context, or way too much of the wrong kind.
I’ve already written about the human half of this equation. Marketing teams need a Context Owner – a curator, an interviewer, the guardian of the single source of truth. That essay was the job description. This series is about building their desk.
Working title: Enterprise Shared Context (ESC).
Yes, it’s a slightly pompous label, but it’s a deeply practical idea. The goal is simple: point a system at the chaotic files you already have, understand what’s inside them, cleanly slice up who gets to see what, and hand that perfectly permissioned slice to Claude, or whatever agent ships next week.
Most modern “knowledge bases” are just file dumps with a search bar. Agents don’t need a dump. They need something smaller, current, and strictly permissioned.
Something to poke
If you want something to poke while I build: [SHIELD](https://github.com/ivans3m/shield) – a fictional spy-agency dump, mixed files, no required layout. Clone it, wander, point an agent at it.
My last project on Cursor
This series is also the last project I’m building inside Cursor as main AI coding agent. I still think it is the best IDE on the market, and a pretty good agent harness. Then came the SpaceX acquisition, and the pricing and usage-cap conditions that followed – which, honestly, never felt great even before the deal closed. So I’m stepping aside for a while: trying Google Antigravity for a bit, and then, most likely, moving over to Anthropic’s tooling. The desk still has to get built. The editor is negotiable.
The roadmap: four parts, five posts
1. This one. The what and the why.
2. Indexing the pile. We’ll look at three approaches, but only build two:
- Simple index (on by default). Filenames, titles, crisp descriptions, and access control.
- LLM wiki (the Karpathy pattern). A compiled, intelligent map sitting on top of raw files. The Context Owner can flip this on for a specific team or the whole company.
- RAG / vector database. Embed everything, retrieve chunks, and pray the walls hold. We looked into this, and we’re skipping it. Access control and vector search are mortal enemies at agency scale.
3. Walls. Everyone gets an index slice that matches their security clearance. Uploads must respect that slice. Pretty wikis tend to leak secrets in paragraphs, not paths. We fix that.
4. Agents. Rules, then CLI, then MCP. Same permission wall, louder tools.
5. The app. What the Context Owner actually clicks when they sit at the desk.
The part that is going to hurt
Ten thousand filenames do not make a memory. A wiki will inevitably cross-link two entirely different clients in one paragraph and proudly call it an “insight.” A vector database will confidently hand you a catering invoice when you asked for a marketing case study.
Building this isn’t easy, but I did the homework. I’m not locking this away in a cathedral.
Next post, we crack open the filing cabinet. Stay tuned.