memboot 0.4.1

Zero-infrastructure persistent memory for any LLM

memboot

Persistent memory for LLMs that works offline. No API keys. No servers. No cloud.

Every other memory tool requires an LLM call to store and retrieve, a managed API, or an HTTP server running in the background. memboot requires none of that. It's SQLite + TF-IDF on your local disk — works on an airplane, works in CI, works without giving a third party access to your codebase.

$ memboot init .
  Indexed 142 files, 1,847 chunks

$ memboot query "authentication flow"
  src/auth/jwt.py:create_token     0.89  "Creates signed JWT with user claims..."
  src/auth/middleware.py:verify     0.84  "Extracts and validates bearer token..."
  src/models/user.py:User          0.71  "User model with hashed password..."

$ memboot remember "Use JWT for API auth, sessions for web" --type decision
  Stored decision #14

$ memboot context "auth" --max-tokens 4000
  # Context: auth (3,842 tokens)
  ## src/auth/jwt.py
  ...

Features

• Smart chunking — AST-aware Python extraction, Markdown heading splits, YAML/JSON key-level, sliding window fallback
• Fully offline — Built-in TF-IDF embeddings with zero external dependencies. Optional sentence-transformers for semantic search
• Episodic memory — Store decisions, patterns, observations alongside your code index
• Context builder — Token-budgeted markdown blocks ready for LLM prompts
• MCP server — Expose memory as tools for Claude Code, Cursor, Claude Desktop, Windsurf, and Zed. memboot install --all-detected wires it into every client you have, zero config
• Zero-config install — One command writes the MCP entry into each client's config with automatic backup
• File ingestion — Ingest external files, PDFs, and web pages into project memory

Install

pip install memboot

Optional extras:

pip install memboot[embed]  # sentence-transformers for semantic embeddings
pip install memboot[mcp]    # MCP server support
pip install memboot[pdf]    # PDF ingestion
pip install memboot[watch]  # File watching for auto-reindex
pip install memboot[web]    # Web page ingestion

Quick Start

# Index a project
memboot init /path/to/your/project

# Search for relevant code and memories
memboot query "authentication flow" --project /path/to/your/project

# Store a decision
memboot remember "Use JWT for API auth, sessions for web" --type decision --project /path/to/your/project

# Get formatted context for an LLM prompt
memboot context "database schema" --project /path/to/your/project --max-tokens 4000

# Check license status
memboot status

CLI Commands

Command	Description
memboot init	Scan, chunk, embed, and index a project
memboot query	Search project memory by similarity
memboot remember	Store an episodic memory (decision, note, observation, pattern)
memboot context	Export a formatted context block with token budget
memboot status	Show license tier and available features
memboot reset	Clear all indexed data and memories
memboot ingest	Add external files, PDFs, or URLs to memory
memboot watch	Watch project and auto-reindex on changes (Pro)
memboot serve	Start MCP stdio server
memboot install	Register memboot as an MCP server in Claude Code / Cursor / Claude Desktop / Windsurf / Zed
memboot uninstall	Remove memboot from a client's MCP config
memboot installed	List which MCP clients currently have memboot registered

How It Works

Project Files ──→ Chunker ──→ Embedder ──→ SQLite Store
                   (AST)      (TF-IDF)     (~/.memboot/)
                                                │
Query Text ─────→ Embedder ──→ Cosine Sim ──→ Results
                                                │
Memories ───────→ Embedder ──→ Store ───────→ Searchable

1. Index — Recursively discover files, chunk by language (Python AST, Markdown headers, etc.), embed with TF-IDF, store in SQLite
2. Query — Embed your query, compute cosine similarity against all chunks and memories, return top-K
3. Remember — Store episodic memories (decisions, patterns, observations) with embeddings for later retrieval
4. Context — Build token-budgeted markdown blocks with source attribution for LLM consumption

Each project gets its own SQLite database at ~/.memboot/{hash}.db. No servers, no API keys, no network calls.

How It's Different

	Works offline	No API keys	No background server	CLI-native
memboot	Yes	Yes	Yes	Yes
Mem0	No (requires LLM)	No	No	No
Memori	No (managed API)	No	N/A	No
OpenMemory	No (HTTP server)	No	No	Partial

Architecture

src/memboot/
├── models.py        # Pydantic v2 data models
├── store.py         # SQLite WAL backend (numpy BLOB serialization)
├── chunker.py       # Language-aware chunking (Python/MD/YAML/JSON/window)
├── embedder.py      # TF-IDF (built-in) + sentence-transformers (optional)
├── indexer.py       # Discovery → chunk → embed → store pipeline
├── query.py         # Cosine similarity search
├── memory.py        # Episodic memory CRUD
├── context.py       # Token-budgeted context builder
├── licensing.py     # Free/Pro tier management
├── cli.py           # 8 Typer CLI commands
├── mcp_server.py    # MCP stdio server (3 tools)
└── ingest/          # External file/PDF/web ingestion

Community

Discord — Join the community

License

MIT