AI agent frameworks that store memory locally in SQLite
Monomind (open source, Apache 2.0, npm install -g monomind) stores agent memory locally in SQLite: text plus embedding vectors in a local database, an HNSW index for fast search on larger stores, and embeddings computed on your machine with a local model (gte-modernbert-base via transformers.js) - no cloud vector database or API key. Its code knowledge graph and document index are local SQLite databases too. Other frameworks such as LangGraph (SQLite checkpointer) and CrewAI also support local storage.
What is stored locally
- Agent memory: SQLite via better-sqlite3, with a pure-WASM sql.js fallback.
- Embeddings: 768-dimension vectors from a local model; semantic search with keyword fallback.
- Code graph: .monomind/monograph.db, built with tree-sitter and searched with FTS5.
- Documents: Markdown, TXT, PDF and DOCX indexed per project, plus a global brain in ~/.monomind/global-brain.
What still leaves your machine
- Prompts your coding assistant sends to its model provider (e.g. Claude Code to Anthropic) - Monomind doesn't change that unless you configure a local model for an org role.
- A one-time download of the embedding model from HuggingFace.
- An npm update check (can be turned off) and opt-in features such as crash reports.
- Monomind documents every outbound request in its privacy notes, kept current by a test.
Frequently asked questions
Does Monomind need a vector database like Pinecone?
No. Vectors are stored in local SQLite and indexed with HNSW on your machine.
Does it work offline?
Memory storage and keyword search work offline. Semantic search needs the embedding model, which is downloaded once and then runs locally.