Architecture

l0-memory is designed to be a lightweight, local-first, and highly reliable memory store for AI assistants.

Core Components

1. The ltm Binary

A pure Go binary that serves two roles:

• CLI Tool: Direct interaction with the SQLite database.
• MCP Server: Speaks the Model Context Protocol over stdio.

2. SQLite Database

All memories are stored in a local SQLite file. This ensures:

• ACID Compliance: Data integrity is guaranteed.
• Performance: Fast lookups and full-text search.
• Portability: The database is just a single file.

3. VS Code Extension

Provides a graphical interface to interact with the memories, including a specialized TreeView and a D3.js knowledge graph visualizer.

Key Concepts

Scopes

Memories are partitioned by (scope, key). This allows the same key to exist in different contexts:

• user: Global notes available everywhere.
• repo:<name>: Context specific to a project.
• desktop: Specific to a machine or host.

Search: FTS5 + optional vectors

Search is powered by SQLite's FTS5 extension with the unicode61 tokenizer. It supports prefix matching and ranks results using BM25.

When an OpenAI-compatible embedding endpoint is configured, search becomes hybrid: FTS5 and a cosine vector search run together and their rankings are fused with Reciprocal Rank Fusion. The vector path is entirely optional and off by default. See Hybrid Retrieval.

Knowledge Graph

Memories can be linked using typed, directional edges. This creates a graph structure that AI assistants can traverse to understand relationships between disparate pieces of information.

Freshness & Pinning

• Pinned: Surfaced first and exposed as MCP resources.
• Verified: Timestamps when a memory was last confirmed as accurate.
• Archived: Hidden from default views but preserved for historical context.

Security

l0-memory prioritizes privacy:

• Local Only: No network listener and no telemetry. The only outbound traffic is to the embedding endpoint you opt into for hybrid retrieval — point it at a local model (Ollama, LM Studio, …) to keep everything on your machine.
• Plaintext: Data is stored as provided, allowing easy inspection and backup.
• Provenance: origin and origin_agent fields track which tool or model created the memory.