LLM Wiki — Synaptic Lattice

SwarmVault

Properties1
tagsentity, tool, llm-wiki, knowledge-graph, local-first

Leitura de 4 min

Overview

SwarmVault is a local-first LLM Wiki & knowledge-graph service (swarmvault.ai), listed in the vault’s Resources page. It combines the wiki pattern with an explicit local knowledge graph.

The 2026-08-02 README capture (swarmclawai/swarmvault) shows it now ships as a real toolchain: an npm CLI (@swarmvaultai/cli, Node ≥24) plus a desktop app for macOS/Windows/Linux that bundles its own runtime. The README positions it as “the local-first LLM Wiki, knowledge graph builder, and RAG knowledge base for AI agents” — a production-grade version of Karpathy’s LLM Wiki gist that turns docs, code, transcripts, notes, and URLs into a durable markdown wiki plus an inspectable, queryable local graph.

Key Facts

  • Type: local-first LLM Wiki + knowledge graph service
  • Positioning: wiki + graph in one local package
  • Listed under Tools & Repositories in Resources
  • Distribution: npm install -g @swarmvaultai/cli (Node ≥24) or the desktop app (macOS/Windows/Linux, no Node required). swarmvault quickstart is the beginner alias for scan: initialize a vault, ingest a local file, directory, or public GitHub repo, compile the wiki and graph, write share artifacts, and open the local graph viewer. No API keys required for the first run.
  • Three-layer architecture explicitly “following the pattern described by Andrej Karpathy”: raw/ (immutable copies of ingested material), wiki/ (LLM-generated + human-authored markdown — “the persistent, compounding artifact”), and swarmvault.schema.md (defines wiki structure and conventions; the user and LLM co-evolve it over time).
  • Typed knowledge graph: every edge tagged extracted, inferred, or ambiguous; automatic contradiction detection (lint --conflicts); compile --approve stages all changes into reviewable approval bundles; new concepts land in wiki/candidates/ first. On disk: state/graph.json (machine-readable graph) + state/retrieval/ (local search index).
  • Hybrid search + offline-first: SQLite full-text merged with semantic embeddings, so queries work without fitting every page into context; the built-in heuristic provider runs locally and offline; optional sharper extraction via a free local LLM (Ollama); cloud providers optional. 30+ input formats, code-aware via tree-sitter AST.
  • Ops surface: swarmvault next (read-only state advisor telling you whether to initialize, ingest, compile, query, review, or refresh), query, graph serve/query/path/explain/callers, doctor, candidate list; Git-backed workflows (--commit), watch mode with git hooks, scheduled automation, an MCP server, and Neo4j/graph export.

Features

Core Architecture

  • Three-layer architecture — raw/ (immutable) → wiki/ (compounding) → schema.md (co-evolved)
  • Local-first — no cloud dependency, runs on your machine
  • npm CLI + desktop app — two distribution modes
  • Node ≥24 — modern runtime requirement

Ingest & Processing

  • 30+ input formats — documents, code, transcripts, notes, URLs
  • Code-aware — tree-sitter AST for code ingestion
  • Public GitHub repo ingest — scan entire repos
  • swarmvault quickstart — one-command init + ingest + compile

Search & Retrieval

  • Hybrid search — SQLite FTS + semantic embeddings
  • Offline-first — built-in heuristic provider, no API keys needed
  • Optional Ollama — free local LLM for sharper extraction
  • Optional cloud — cloud providers available but not required

Knowledge Graph

  • Typed edges — every edge tagged: extracted, inferred, or ambiguous
  • Contradiction detection — lint --conflicts finds conflicting claims
  • Approval bundles — compile --approve stages changes for review
  • Candidates — new concepts land in wiki/candidates/ first
  • Graph visualization — local graph viewer
  • Neo4j export — for advanced graph queries

MCP / Tools

  • MCP server — Model Context Protocol integration
  • swarmvault next — read-only state advisor
  • swarmvault doctor — health diagnostics
  • Graph commands — serve, query, path, explain, callers

Quality & Governance

  • lint --conflicts — automatic contradiction detection
  • Approval workflow — review before committing changes
  • Git-backed — full version history with git hooks
  • Watch mode — automated re-ingest on file changes

Integration

  • Claude Code, Codex — agent-compatible
  • Obsidian — wiki works as Obsidian vault
  • Neo4j — graph export for advanced queries
  • Scheduled automation — cron-like scheduling

Data & Storage

  • Markdown files — wiki output in markdown
  • state/graph.json — machine-readable graph
  • state/retrieval/ — local search index
  • SQLite — full-text search database

Relationships

  • “uses” knowledge-graph — explicit graph is core to it
  • “belongs-to” llm-wiki — local-first implementation of the Karpathy pattern
  • “sibling-of” cognee — sibling graph-backed local option
  • “compatible-with” open-knowledge-format — local-first aligns with OKF portability
  • “inspired” memex — README frames itself in the Vannevar Bush Memex (1945) tradition: connections between sources as valuable as the sources
  • “references” andrej-karpathy — three-layer architecture explicitly follows his LLM Wiki pattern

Implications

SwarmVault is a local-first graph+wiki option — relevant if the user wants the graph layer (Cognee-style) without going to the cloud.

The 2026-08-02 README capture upgrades this from a bare listing to a concrete, implementable design. It validates the exact choices this vault already made (immutable raw/, compounding wiki/, co-evolved schema, typed relationships, approval queues, contradiction handling, forward-only linking) and adds two features this vault does not yet ship: an interactive graph viewer and a copyable/visual share-kit. Its “From Gist to Production” comparison table is also a useful maturity checklist when evaluating other LLM Wiki implementations.

Sources

^[wiki/concepts/resources.md]

Visão de gráfico

Backlinks