Synto

Turn your raw notes into a self-improving, interlinked wiki — powered by a local LLM.

You drop Markdown notes in a folder. Synto reads them with a local LLM, extracts the concepts they contain, and writes one cross-linked article per concept. Every note you add makes the wiki richer. Every article stays on your machine unless you decide otherwise.

After setup (~5 minutes) you have: a structured wiki built from your notes, a queryable knowledge base that works without embeddings or a vector database, and an agent-ready pack that Claude, Cursor, or any file-aware AI can install and reason over — including reading your sources' exact words on demand over MCP.

The idea

Andrej Karpathy described it as the LLM Wiki: a personal knowledge base where the model doesn't just store what you tell it — it synthesizes, cross-references, and keeps everything current. You add raw material; it does the bookkeeping.

The key insight: treat your notes as source material, not as the final artifact. A raw note is a claim about the world. A wiki article is the compiled, cross-linked explanation of a concept derived from many notes. The LLM does the compilation step.

You write raw notes  →  LLM extracts concepts  →  Wiki articles grow  →  Agent-ready pack
      raw/                    (automatic)              wiki/             .synto/exports/
  quantum.md           "Qubit", "Superposition"     Qubit.md
  ml-basics.md         "Neural Net", "SGD"          Superposition.md ←── [[wikilinks]]
  physics.md           "Qubit"  ← same concept      Neural_Network.md
                              │
                     Duplicates merge, not multiply.
                     One article per concept, fed by all relevant notes.

Unlike a chatbot that forgets, the wiki persists and compounds. Every note you add and every draft you review makes it smarter.

How it works

Four stages. Two LLM tiers.

┌────────────────────────────────────────────────────────────────────────┐
│  Stage 1: Import  Stage 2: Ingest  Stage 3: Compile  Stage 4: Export  │
│                                                                        │
│  synto add ─────┐                                                      │
│  (PDF/md/txt)   ├─ raw/*.md ──────► wiki/.drafts/ ──────► exports/    │
│  .synto/sources/┘  fast model       heavy model      agent-ready      │
│                    (4B params)      (14B+ params)     directory        │
│                                                                        │
│  archives:         extracts:        writes:           produces:       │
│  · original file   · concepts       · one article     · INDEX.json    │
│  · segments        · summaries        per concept     · AGENTS.md     │
│  · source type     · relationships  · cross-linked    · CLAUDE.md     │
│                    · language         [[wikilinks]]   · articles/     │
│                                     · source-type                     │
│                                       prompt                          │
└────────────────────────────────────────────────────────────────────────┘

Why two LLM tiers? Analysis is pattern-matching — a 4B model running locally can extract "this note is about Qubit, Superposition, and Entanglement" reliably and fast. Writing a coherent, cross-linked article requires more reasoning — a 14B+ model does this well. Splitting the work keeps the pipeline cheap and fast on consumer hardware.

Your vault layout after synto init: