Smriti

Local-first persistent memory for AI agents via MCP.

Smriti (स्मृति) — Sanskrit for "memory, remembrance"

One brain. Every agent. Zero cloud. Zero cost.

What is this?

A standalone MCP server backed by sqlite-vec and a local embedding model. Install it, point any MCP-compatible agent at it, and every AI you use shares one persistent, semantically searchable memory.

	Smriti	Cloud alternatives
Setup	npx smriti	Accounts + API keys + config
Cost	$0	Variable
Privacy	100% local	Data on external servers
Offline	Full functionality	Needs internet
Portability	Single .db file	DB export/migration

Install

npm install -g smriti

Usage

# stdio mode (for Claude Code, Cursor, etc.)
smriti

# HTTP mode (for remote agents)
smriti --http --port 3838

MCP Client Configuration

Claude Code

Add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "memory": {
      "command": "smriti"
    }
  }
}

Cursor

Add to MCP settings:

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["smriti"]
    }
  }
}

Any MCP client (HTTP mode)

smriti --http --port 3838

Then point your client at http://localhost:3838/mcp.

Tools

Tool	Description
capture	Store a thought with auto-extracted metadata
search	Semantic search — find thoughts by meaning
recall	Browse recent memories with filters
forget	Delete a specific memory by ID
context	Get structured context bundle for a topic
stats	Memory patterns and insights

Resources

URI	Description
memory://recent	Last 24h of thoughts
memory://topics	Topic index with counts
memory://people	People mentioned + context
memory://stats	Overall memory statistics

Prompts

Name	Description
brain-dump	Guided capture session
weekly-review	End-of-week synthesis
migrate	Import memories from other sources

Configuration

Config lives at ~/.smriti/config.json:

{
  "db_path": "~/.smriti/brain.db",
  "embedding": {
    "provider": "onnx",
    "model": "Xenova/all-MiniLM-L6-v2"
  },
  "extraction": {
    "provider": "rules"
  },
  "server": {
    "transport": "stdio",
    "port": 3838
  }
}

How it works

1. You (or an agent) call capture with text
2. Smriti generates a vector embedding locally (all-MiniLM-L6-v2 via ONNX)
3. Regex-based extraction pulls out people, topics, actions, and classifies the type
4. Everything is stored in a single SQLite file with sqlite-vec for vector search
5. search finds thoughts by semantic similarity, not just keywords
6. All data stays on your machine — nothing leaves localhost

License

MIT