Using an Obsidian Vault as Your Agent's Knowledge Base

Why a vault is already the right shape

Most attempts to give an agent knowledge start by building a corpus. If you keep a vault, you have been building one for years without calling it that.

The properties that matter are already there. The files are plain text, so no export or connector stands between the agent and the content. They are local, so you decide what leaves your machine. They are written in your vocabulary, which is what makes retrieval match the way you will ask. They are linked, which encodes relationships nobody would reconstruct manually. And they are usually under version control or at least backed up, so the history exists.

Compare that with the usual alternative, which is knowledge spread across a wiki, a ticket system, three chat tools, and several people's memories. The vault has already solved the hardest part, which is that the material exists in one place in a machine readable form.

What is missing is one capability: finding the right passage for a question, without loading everything.

The gap: retrieval, not storage

Two approaches fail here, and it is worth being precise about why.

Loading the vault into context. Impossible past a small vault, and counterproductive before that. Even where it fits, filling the window with mostly irrelevant notes degrades attention on the parts that matter and costs a fortune on every request.

Grepping the vault. Better than nothing and structurally limited. Grep finds the string you typed. Your note about deployment might say shipping, release, or rollout, and an agent asking about deployment finds nothing. Keyword search fails exactly when your vocabulary drifted, which in a vault written over years is constantly.

What works is retrieval by meaning: the agent asks a question, the relevant passages come back, and only those enter the window.

That is what RDK does with a vault. Files from local vaults, docs, and code are indexed as encrypted private chunks, and agents search those chunks before querying a model. Token spend drops 80 to 90 percent on repeated or reference-heavy work because the answer is retrieved instead of regenerated, and because the chunks are encrypted, the content stays unreadable to anyone else while remaining searchable by your agents.

Stacked retrieval sets the shape in practice: your private vault answers 40 to 65 percent of queries, the public network adds 15 to 20 percent for general knowledge someone else documented well, and the model handles the remaining 5 to 10 percent.

Curation is the part that decides quality

An index over everything you have ever written will return something for every query. That is the problem rather than the goal.

Vaults accumulate drafts, half formed ideas, notes from a project that ended, opinions you have since changed, and clippings you never read. Retrieved confidently, superseded material is worse than missing material, because it arrives with the same authority as your current thinking.

Three rules keep an index useful.

Exclude by folder. Daily notes, inbox, clippings, and anything you would describe as scratch. If you would not hand it to a colleague as an answer, do not let an agent hand it to you as one.

Mark superseded notes rather than deleting them. A front matter field is enough. You keep the history and stop retrieving conclusions you have abandoned.

Use links as a quality signal. A note that many others reference is usually a settled conclusion. A note nothing links to is often a passing thought, and weighting retrieval toward the connected parts of a vault matches how the vault is actually organized.

This is judgment work that takes an afternoon and outperforms every tuning knob available to you afterwards.

What this changes day to day

The visible change is that you stop re-explaining yourself. Project background, past decisions, the conventions you follow, and the reasoning behind an approach are retrieved rather than pasted, which removes both the token cost and the failure mode where you forget to include one of the three paragraphs.

The less obvious change is that the vault stops being write-only. Most people write more notes than they ever read again, and the marginal value of a note is close to zero if nothing surfaces it at the right moment. Making the vault retrievable turns years of writing into something that pays out repeatedly, which is also the strongest argument for keeping notes at all.

The discipline that keeps it working is small: write conclusions with their reasoning rather than conclusions alone, since a decision without its constraint gets misapplied later by an agent as easily as by a person.