A local-first knowledge vault that makes your LLM do the filing

What it does

SwarmVault ingests files, code, URLs, and transcripts into an immutable raw/ store, then compiles them into a structured markdown wiki and a typed knowledge graph that lives entirely on your disk. You get hybrid search, an interactive graph viewer, and reviewable approval queues to sanity-check LLM-generated pages before they become canon. It runs offline out of the box; cloud or local LLM providers are strictly optional upgrades.

The interesting bit

The project treats Karpathy’s LLM Wiki gist as a spec and builds the missing production toolchain around it: contradiction detection, candidate staging for new concepts, and a co-evolvable swarmvault.schema.md that acts as a contract between you and the model. The genuinely useful drudgery is the maintenance automation—every graph edge is tagged extracted, inferred, or ambiguous, and the vault can lint for conflicts so bad inferences do not compound silently.

Key highlights

• Fully offline first run via a built-in heuristic provider; Ollama and cloud models are optional.
• Hybrid search that merges SQLite full-text search with semantic embeddings, so you can query without stuffing the entire vault into context.
• Code-aware ingestion via tree-sitter AST parsing, plus support for 30+ input formats including transcripts, papers, and mail exports.
• Agent-facing outputs: context packs, task ledgers, an MCP server, and export to Neo4j or Obsidian.
• Git-friendly workflows with watch mode, scheduled automation, and reviewable compile --approve bundles.

Verdict

Developers drowning in scattered codebases, research papers, and meeting transcripts who want a durable, queryable archive without shipping everything to a cloud API should look here. If you just want a simple note-taking app and have no patience for schemas, approval queues, or graph exports, this is overkill.