On the art of remembering

The case for
compiled knowledge

Knowledge Base is built on a simple premise: the LLM should do the bookkeeping that humans abandon after two weeks. It’s not a note-taking app. It’s a compilation system.

history_edu

The Origin

On April 3, 2026, Andrej Karpathy published a pattern that resonated with 12 million people: place an LLM between your feed and your memory. Let it extract concepts, cross-reference sources, and maintain a persistent wiki that compounds over time.

The idea exploded because it named a problem everyone felt but nobody had solved: we consume more than ever and retain almost nothing. The average knowledge worker reads 50+ articles a month and remembers five. The Ebbinghaus forgetting curve is merciless — 70% of what you learn today will be gone by tomorrow.

Karpathy’s insight was deeper than “just save your bookmarks.” He described a system where knowledge is compiled, not stored. Where cross-references emerge automatically. Where every new source enriches everything that came before. Where contradictions between papers are flagged, not buried. Where asking a question makes the entire system richer.

He called it “room for an incredible new product.” We’re building that product.

trending_down

The Knowledge Retention Crisis

We read more than any generation in history. But reading without compiling is just entertainment. Knowledge tools have failed because they ask youto do the work — tagging, linking, organizing. Nobody does that after the first two weeks.

50+

Articles read per month

70%

Forgotten within 24 hours

Actually remembered

Cross-referenced

architecture

Compilation, Not Storage

Most knowledge tools are filing cabinets. You put stuff in, maybe tag it, and hope you can find it later. Knowledge Base is fundamentally different. Here are the principles we build on:

auto_awesome

Compiled, not stored

Knowledge is compiled once and kept current, not re-derived on every query. Your wiki is a persistent, growing artifact.

hub

Cross-referenced by default

Cross-references are already there, contradictions already flagged. You never have to manually link ideas again.

person

Human curates, LLM maintains

You choose what to learn. The LLM handles summarizing, cross-referencing, and bookkeeping that nobody does after week two.

replay

Every question compounds

Good query answers get filed back into the wiki as new pages. Every question you ask makes the entire system richer.

lock

Yours to keep

Standard Markdown files, Obsidian-compatible, exportable at any time. No lock-in. Your knowledge, your format.

memory

Retention, not consumption

Spaced repetition, reading streaks, and knowledge graphs help you actually remember what you read.

How It Works

Four steps from raw content to compiled knowledge. The LLM does the heavy lifting so you can focus on learning.

cyclone

Ingest

Connect 13+ sources. Every bookmark, highlight, and save is pulled into your private substrate.

account_tree

Compile

LLM extracts concepts, entities, and relationships. Cross-references with your existing knowledge.

auto_stories

Wiki

A persistent, interlinked wiki in standard Markdown. Obsidian-compatible. Exportable. Yours.

psychology

Query

Ask anything. Get synthesized answers with citations to your own sources. Filed back as wiki pages.

code

Under the Hood

Knowledge Base is a three-layer architecture: raw sources (immutable input), compiled wiki (LLM-generated, interlinked Markdown), and schema (conventions, workflows, structure).

The backend is Python with a SQLite database, FTS5 full-text search, and Pydantic models. The LLM layer is provider-agnostic via LiteLLM — use OpenAI, Anthropic, or run fully local with Ollama. The frontend is Next.js with React 19, Tailwind CSS, and D3.js for knowledge graphs.

The wiki output is standard Markdown with YAML frontmatter and[[wikilinks]]— fully compatible with Obsidian, Logseq, or any Markdown editor. No lock-in, ever.

13+

Source connectors

100%

Obsidian compatible

Local

LLM option

Open

Markdown format

group

The Team