Dossier | Twofold

What it does

OKF: the Open Knowledge Format

One concept per file — a decision, a process, a term — as Markdown + YAML. Typed, sourced, and scored. cat-able, diff-able, git-clone-able.

Your git is the system of record

The repo is the single source of truth, not an export of it. Vector and graph indexes are derived caches — disposable and rebuildable at any time.

Served over GraphRAG and MCP

Agents and tools read the same memory through Model Context Protocol and graph-aware retrieval — one substrate for Claude Code, CI, and internal apps.

Humans curate, agents extend

Agents propose candidate atoms after every piece of work; your people approve what becomes permanent. Curation is a first-class step, not an afterthought.

Graph and board, live over the files

The graph renders the knowledge base as a living web — every atom a node, every typed edge a relationship. The board is an editorial kanban over the task atoms that ARE your backlog.

What it is

Every organization is generating more knowledge than ever — and losing most of it between sessions, tools, and departures. Dossier is a sovereign learning loop: your institutional knowledge as plain-text files in your own git repository, curated by your people, extended by agents after every piece of work.

Nothing proprietary sits between you and your own memory. It diffs, reviews, branches, and rolls back like any other repository — because it is one.

How the loop works

Four stages, compounding:

Ingest. Point Dossier at files, sites, and docs.
Extract. Content becomes atomic OKF files plus typed edges — a knowledge graph committed to your git repository.
Serve. The graph is queryable over GraphRAG and exposed over MCP, so every agent and tool reads from the same memory.
Curate. Humans curate, agents extend. Each unit of work proposes candidate updates; your people decide what becomes permanent.

More work in, better memory out. That is the compounding part.

OKF, the Open Knowledge Format

The substrate is deliberately boring: Markdown + YAML, one concept per file. An atom is shaped like this:

---
type: decision
title: Billing store runs on Postgres
confidence: verified
source: adr/0042-billing-store.md
---

We chose Postgres over a document store because billing
queries are relational and audit requirements demand
transactional history.

Every atom is typed, cites its source, and carries a confidence score — verified, asserted, or inferred — so a consumer knows the difference between what the organization has confirmed and what an agent guessed.

OKF is not an export of your knowledge. It is the system of record itself. Vector and graph indexes are derived caches: disposable, rebuildable, never load-bearing.

Two live views

The graph renders the whole knowledge base as a living web — every atom a node, every typed edge a relationship.
The board is an editorial kanban over the task atoms that are your backlog. Not a mirror of your tracker — the files themselves, arranged by state.

Both views read and write the same files in the same repo.

Why files in git

Rented memory dies with the subscription. A vector store you can’t inspect isn’t institutional memory; it’s a liability with an API. Dossier’s bet is that the boring substrate is the right one: files you can read, history you can audit, and a review step that keeps humans in charge of what the organization believes.

Status

Private beta. The first build is shipping to early adopters now. See the live product at dak-dossier.vercel.app, or get in touch if institutional memory is a problem you’re tired of having.

FAQ

What is OKF?

The Open Knowledge Format: one concept per plain-text file — a decision, a process, a term — as Markdown with YAML frontmatter. Each atom is typed, cites its source, and carries a confidence score: verified, asserted, or inferred. Because atoms are files, they cat, diff, and git clone like everything else you ship.

Is this another vector database?

No. In Dossier, indexes are caches. Vector and graph indexes are derived from the files and can be thrown away and rebuilt at any time. The git repository is the system of record — inspectable, reviewable, and portable.

How do agents actually use it?

Dossier serves the knowledge graph over GraphRAG and MCP. Any MCP-capable agent — Claude Code included — reads atoms for context before work and proposes candidate atoms after. Humans review the proposals; only approved atoms enter the record.

What keeps the memory trustworthy?

Two mechanisms. Every atom is scored — verified, asserted, or inferred — so consumers know how much weight a claim carries. And curation is gated: agents extend, humans approve. Nothing becomes institutional truth without a person signing off.

What happens if we stop using Dossier?

You keep everything. The knowledge base is plain files in your own repository — never locked to us, to Google, or to any model. Walk away and the repo still works with any tool that can read files.