MemG TypeScript SDK

Memory layer for LLM applications. Zero runtime dependencies.

Install

npm install memg

Quick Start

Proxy Mode (Recommended)

Route LLM traffic through the MemG proxy for transparent memory recall and extraction.

import { MemG } from 'memg';
import OpenAI from 'openai';

const openai = MemG.wrap(new OpenAI(), { entity: 'user-123' });

const response = await openai.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'What do I usually order?' }],
});

Client Mode

Intercepts calls locally without running the proxy. Requires only the MCP server.

import { MemG } from 'memg';
import Anthropic from '@anthropic-ai/sdk';

const anthropic = MemG.wrap(new Anthropic(), {
  entity: 'user-123',
  mode: 'client',
});

const response = await anthropic.messages.create({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Remind me about my preferences' }],
});

Direct Memory Operations

import { MemG } from 'memg';

const m = new MemG();

// Add memories (flexible input)
await m.add('user-123', 'likes coffee');
await m.add('user-123', ['works at Acme', 'prefers dark mode']);
await m.add('user-123', [
  { content: 'allergic to peanuts', type: 'identity', significance: 'high' },
]);

// Search
const results = await m.search('user-123', 'food preferences');
console.log(results.memories);

// List all
const all = await m.list('user-123', { type: 'identity' });

// Delete
await m.delete('user-123', 'memory-uuid');
await m.deleteAll('user-123');

Chat (Session-Aware)

chat() manages the full memory loop: sessions, history injection, recall, LLM call, exchange persistence, and extraction.

import { MemG } from 'memg';

const m = new MemG({ openaiApiKey: process.env.OPENAI_API_KEY, embedProvider: 'openai' });
await m.init();

const res = await m.chat(
  [{ role: 'user', content: 'I just moved to Seattle' }],
  'user-123'
);
console.log(res.content);

// Follow-ups are history-aware — the session tracks prior turns.
const res2 = await m.chat(
  [{ role: 'user', content: 'What city did I mention?' }],
  'user-123'
);

Low-Level MCP Client

import { MemGClient } from 'memg';

const client = new MemGClient('http://localhost:8686');

await client.add('user-1', [{ content: 'likes TypeScript' }]);
const results = await client.search('user-1', 'programming languages');

Gemini

import { MemG } from 'memg';
import { GoogleGenerativeAI } from '@google/generative-ai';

const genai = new GoogleGenerativeAI(process.env.GEMINI_API_KEY!);
const model = MemG.wrap(genai.getGenerativeModel({ model: 'gemini-2.5-flash' }), {
  entity: 'user-123',
  nativeConfig: { geminiApiKey: process.env.GEMINI_API_KEY, embedProvider: 'gemini' },
});

const result = await model.generateContent('What do you remember about me?');

Custom Store (MySQL, Postgres, etc.)

The SDK defaults to SQLite via better-sqlite3, but you can pass any object implementing the Store interface:

import { MemG, type Store } from 'memg';

class MyPostgresStore implements Store {
  // Implement all interface methods...
}

const m = new MemG({ store: new MyPostgresStore(connectionString) });

See Store in store.ts for the full interface contract (28 methods).

Supported Providers

| Provider | wrap() | Extraction | Embeddings | |----------|----------|------------|------------| | OpenAI | new OpenAI() | Yes | Yes | | Anthropic | new Anthropic() | Yes | No | | Gemini | genai.getGenerativeModel() | Yes | Yes |

Configuration

| Option | Default | Description | | ---------- | -------------------------- | ------------------------------------ | | mcpUrl | http://localhost:8686 | MCP server URL | | proxyUrl | http://localhost:8787/v1 | MemG proxy URL | | entity | — | Entity identifier for memory scoping | | mode | native | native, proxy, or client | | extract | true | Extract knowledge from responses | | nativeConfig.store | — | Custom Store implementation | | nativeConfig.embedProvider | sentence-transformers | sentence-transformers, openai, or gemini | | nativeConfig.geminiApiKey | — | Gemini API key for embeddings/LLM |

Requirements

• Node.js >= 18 (native fetch)
• For native mode: better-sqlite3 (auto-installed)
• For proxy/client mode: MemG server running
• OpenAI SDK >= 4.0.0 (optional peer dependency)
• Anthropic SDK >= 0.20.0 (optional peer dependency)
• @google/generative-ai (optional peer dependency)