@maskweaver/memory

Simple memory system with embeddings and vector search.

Features

• Text Chunking: Break long documents into overlapping chunks
• SQLite Storage: Fast local storage with FTS5 full-text search
• Vector Search: Cosine similarity search over embeddings
• Hybrid Search: Combine vector and text search
• Multiple Providers: Ollama, OpenAI, Voyage, OpenRouter

Usage

import {
  initDatabase,
  getDbPath,
  indexFile,
  hybridSearch,
} from '@maskweaver/memory';
import { OllamaProvider } from '@maskweaver/memory';

// Initialize database
initDatabase(getDbPath());

// Create embedding provider
const provider = new OllamaProvider({
  model: 'bge-m3',
  baseUrl: 'http://localhost:11434',
  dimensions: 1024,
});

// Index a file
await indexFile(
  './memory/MEMORY.md',
  (text) => provider.embed([text]).then(e => e[0])
);

// Search
const results = await hybridSearch(
  'Rob Pike system programming',
  await provider.embed(['Rob Pike system programming']).then(e => e[0]),
  { limit: 5, minScore: 0.3 }
);

console.log(results);

Architecture

core.ts          # Basic utilities (hash, similarity, paths)
chunking.ts      # Text chunking with overlap
store/sqlite.ts  # SQLite storage with FTS5
search/hybrid.ts # Vector + text hybrid search
indexer.ts       # File indexing pipeline
providers/       # Embedding providers

Design Principles

Following Rob Pike's philosophy:

• Clear is better than clever - Simple, readable code
• Errors are explicit - No silent failures
• Minimal dependencies - Only what we need
• Single responsibility - Each module does one thing well

Dependencies

• better-sqlite3 - Fast SQLite library
• Embedding providers (optional)

License

MIT