Semantic Memory

semantic-memory

Local semantic memory with PGlite + pgvector. Budget Qdrant that runs anywhere Bun runs.

Why

You want semantic search for your AI agents but don't want to run a vector database server. This gives you:

• Zero infrastructure - PGlite is Postgres compiled to WASM, runs in-process
• Real vector search - pgvector with HNSW indexes, not some janky cosine similarity loop
• Collection-based organization - Different collections for different contexts (codebase, research, notes)
• Configurable tool descriptions - The Qdrant MCP pattern: same tool, different behaviors via env vars
• Effect-TS - Proper error handling, resource management, composable services

Install

# npm/bun/pnpm
npm install semantic-memory

# Need Ollama for embeddings
brew install ollama
ollama pull mxbai-embed-large

CLI

# Via npx
npx semantic-memory store "The auth flow uses JWT tokens stored in httpOnly cookies"
npx semantic-memory find "how does authentication work"

# Or install globally
npm install -g semantic-memory
semantic-memory store "React component patterns" --collection code
semantic-memory find "components" --collection code

# Full-text search (no embeddings)
semantic-memory find "JWT" --fts

# Add metadata
semantic-memory store "API rate limits are 100 req/min" --metadata '{"source":"docs","priority":"high"}'

# List, get, delete
semantic-memory list
semantic-memory get 
semantic-memory delete 

# Stats
semantic-memory stats

Collections for Context

Collections let you organize memories by purpose. The collection name carries semantic meaning:

# Codebase analysis - store patterns, architecture notes, API quirks
semantic-memory store "Auth uses httpOnly JWT cookies with 7-day refresh" --collection codebase
semantic-memory store "The useOptimistic hook requires a reducer pattern" --collection codebase
semantic-memory find "authentication" --collection codebase

# Research/learning - concepts, con