braindump

Dump your thoughts as Markdown. Ask questions across all of them at once.

---

TL;DR

# 1. Install
pip install braindump-ai

# 2. Initialize a workspace (creates spikes/, wiki/, and llm.json)
braindump init ~/my-knowledge-base

# 3. Run the local server
braindump run ~/my-knowledge-base

Open http://localhost:8000 in your browser. Write spikes in the editor, ask questions in the query bar, explore the knowledge graph.

Prerequisites: Claude Code credentials in ~/.claude/ (run claude login once to authenticate — the claude binary itself does not need to be on PATH, the SDK ships a bundled copy).

---

The problem

Every knowledge tool forces a choice: organize first, think later. You have to decide where a note belongs before you write it. You end up with isolated notebooks that cannot talk to each other — ask the same question in two places and you get two disconnected answers.

What braindump does differently

braindump treats each Markdown file as a spike — a timestamped, tagged unit of thinking. Write in the built-in editor. The app automatically builds a living wiki that connects related spikes by shared tags, semantic relationships, and temporal proximity. No manual linking. No folders to organize.

When you ask a question, braindump reasons across your entire corpus at once — not just one notebook, not just keyword search. It gives you a grounded answer that cites exactly which spikes it drew from.

The knowledge graph is hierarchical: zoom out to see the major themes in your thinking, zoom in to individual spikes and their connections. The structure emerges from your writing, not from a taxonomy you imposed upfront.

Unique selling point

Cross-corpus AI reasoning over a self-organizing, portable knowledge graph — entirely on your machine.

• Everything is plain .md files. No proprietary format, no lock-in.
• No data leaves your machine. All LLM inference uses the claude CLI and your existing Anthropic subscription.

How it works

Write a spike in the UI → braindump saves it and asynchronously updates a wiki index backed by three plain markdown files:

• wiki/index.md — LLM-authored summaries of every spike
• wiki/connections.md — LLM-authored explicit semantic links between related spikes
• wiki/hierarchy.md — LLM-authored thematic community groupings

Ask a question → the LLM reads the compiled index and answers with inline citations ([1], [2], …) pointing back to the source spikes.

Explore the graph → zoom out to major topic clusters derived from the hierarchy, zoom in to individual spikes and their connections.

Architecture

Wiki layer

The core innovation is the wiki layer: a set of human-readable markdown files maintained by the LLM. Every create/update/delete triggers a background job that rewrites index.md, connections.md, and hierarchy.md in sequence. The wiki doubles as both the retrieval index and a browsable knowledge artifact.

A lightweight meta.json cache (no LLM involved) keeps spike listings fast. A wiki/log.md tracks all events for diagnostics.

Query pipeline

No vector embeddings. One LLM call per query:

1. Feed the hierarchy hierarchy.md and full index.md summaries as context
2. LLM answers with [N] citations; citations are parsed into source cards

Because the index summaries are rich, raw spike content is not sent to the model.

Knowledge graph

Derived purely from the wiki markdown files — no database:

Zoom	View	Source
0	Community clusters only	hierarchy.md
1	Clusters + spike nodes	hierarchy.md
2	Spikes + tag / semantic / temporal edges	meta.json, connections.md

Stack

• Backend: Python >= 3.13 and fastapi
• Frontend: React + Cytoscape.js
• LLVM: Claude (via claude CLI)
• Storage: plain Markdown.