Local Setup ($0 with Ollama)

Run Mnemo entirely on your machine — no API keys, no cloud services, no cost.

Prerequisites

Install Ollama and pull the embedding model:

ollama pull bge-m3

Usage

import { createMnemo } from '@mnemoai/core';

const mnemo = await createMnemo({
  embedding: {
    provider: 'openai-compatible',
    apiKey: 'ollama',                          // any non-empty string works
    baseURL: 'http://localhost:11434/v1',       // Ollama's OpenAI-compatible endpoint
    model: 'bge-m3',
    dimensions: 1024,
  },
  dbPath: './local-memory-db',
});

await mnemo.store({ text: 'User likes hiking on weekends' });
const results = await mnemo.recall('hobbies');

Optional: Local LLM for Smart Extraction

For the full pipeline (smart extraction + contradiction detection), also pull an LLM:

ollama pull qwen3:8b

Then configure:

const mnemo = await createMnemo({
  embedding: {
    provider: 'openai-compatible',
    apiKey: 'ollama',
    baseURL: 'http://localhost:11434/v1',
    model: 'bge-m3',
    dimensions: 1024,
  },
  llm: {
    apiKey: 'ollama',
    baseURL: 'http://localhost:11434/v1',
    model: 'qwen3:8b',
  },
  dbPath: './local-memory-db',
});

Optional: Local Reranker

For cross-encoder reranking without an API:

ollama pull bge-reranker-v2-m3

const mnemo = await createMnemo({
  embedding: { /* ... */ },
  retrieval: {
    rerank: 'cross-encoder',
    rerankProvider: 'ollama',
    rerankModel: 'bge-reranker-v2-m3',
    rerankEndpoint: 'http://localhost:11434',
  },
  dbPath: './local-memory-db',
});

Performance

Component	Model	RAM	Speed
Embedding	bge-m3	~1.1GB	~90ms/query
LLM	qwen3:8b	~5GB	~500ms/extraction
Reranker	bge-reranker-v2-m3	~1GB	~50ms/rerank

All models run on CPU. GPU acceleration available if you have one.