You Probably Don’t Need RAG for Your Agent’s Knowledge Base

We connected 16 repos with markdown files, grep, and zero new infrastructure

The moment your project outgrows a single repo, your coding agents start losing context. They can read the README in front of them, but they don't know how the API contracts connect to the ML pipelines or why the deployment is configured the way it is. We hit this at Lander Analytics with a 16+ repo project and solved it with the simplest thing we could think of: a shared directory of markdown files.

A lot of the most important knowledge — deployment patterns, cross-service contracts, architectural decisions — doesn't belong in any one repo. It sits between repos. Without a shared knowledge base, that information either lives in someone's head or gets duplicated (and inevitably drifts) across multiple READMEs.

How do agents actually use this? Not through retrieval-augmented generation (otherwise known as “RAG”). They use their native tools: grep, glob, and file reads. The agent searches for a term, finds the relevant markdown file, reads it, and has the context it needs. This is agentic search, not RAG. The agent decides what to look for and when.

We built a custom skill that teaches agents how to navigate and update the knowledge base. The skill points the agent to 00-start-here.md and the relevant repo note first, then defines rules for when to update docs — after architectural changes, new cross-repo contracts, or durable decisions. When an agent makes a meaningful change, it can update the corresponding doc in the same workflow. The docs stay current because maintaining them is part of the development loop, not a separate chore.

For the humans on the team, Obsidian gives us a browsable GUI on top of the same markdown files. No syncing, no separate tool, just point it at the directory and you get linked notes, search, and a graph view of how everything connects.

Here's the directory layout:

The numbered prefixes (10-, 20-, 30-) are deliberate. They let an agent browse at a topic level and only pull in the files it needs, rather than dumping the entire knowledge base into its context window. An agent working on the frontend can read 20-repos/frontend.md without also loading the orchestration or infrastructure docs. The 00-start-here.md file acts as an index, pointing agents toward whatever is relevant to their current task.

Andrej Karpathy made a similar observation recently about using LLMs to maintain personal knowledge bases stored as markdown. His takeaway: "I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents." The AGENTS.md standard is gaining traction for the same reason. Markdown is what agents already know how to read.

There are clear limitations. This works at small to medium scale. Once your knowledge base exceeds what an agent can reasonably search and read in a few hops, you'll need a retrieval layer. The structure of your directories matters, and it can drift if nobody is actively maintaining it. For large enterprises with millions of documents, this is not the answer.

But for a team managing a complex multi-repo project? Directories and .md files. No infrastructure required. Start there.

We can help you figure this out

This is the kind of problem we work on at Lander Analytics. We help teams structure knowledge bases for coding agents, set up the skills and workflows to keep documents current, and figure out when simple markdown is enough versus when you need something more. Whether your team is adopting coding agents or trying to get more out of the ones you already use, reach out at info@landeranalytics.com. We’re here to help.

Joe Marlo

Director of Data Science

Lander Analytics

---

Subscribe to our Substack and below to our monthly emails for practical AI strategies for your organization: what to build, what to avoid, and how to make systems reliable in the real world.

Work with us: If you want help identifying the right first workflow, building a permissioned knowledge base, or training your team to ship responsibly, reach out at info@landeranalytics.com.

About the author: Joe Marlo is Director of Data Science at Lander Analytics, where he designs agentic workflows, statistical models, and interactive frontends that put rigorous analysis into production.