Comparison

A Mem0 alternative for one person, not a developer team

Q: Does Mem0 need an OpenAI key?

With its defaults, yes. Memory() uses OpenAI for the language model and embeddings, and in our run on mem0ai 2.2.1 without OPENAI_API_KEY it failed at setup with a missing credentials error. You can configure other providers. In a second run on October 24, 2026, with a local embedder (fastembed), a local Qdrant path and an Ollama language model block, Memory.from_config set up in 0.4 seconds with no OpenAI key, and storing text as given worked. Extracting memories from conversation also needs a running Ollama, and without one it failed with an LLM error.

By Matthew Firth, founder of ModelBrain · Oct 19, 2026 · 5 min read

If you are one person who wants Claude, Cursor or ChatGPT to remember things, ModelBrain asks fewer questions before it saves anything. If you are a developer adding memory to your own app, Mem0 is the tool built for that. Here is our own proposal for comparing memory tools, the first-memory test: count what you must install, supply and decide between "I want a memory" and "one memory is saved."

On the left a thin amber wire passes through six small black gates in a row, on the right a single wire passes through one ring, and both end at a glowing amber capsule

What we ran on October 19, 2026

We installed mem0ai 2.2.1 from PyPI into a clean Python 3.13 environment on Linux. We ran from mem0 import Memory and then Memory() with no OPENAI_API_KEY set. It stopped with an OpenAIError saying credentials were missing, before any memory was saved. We then printed the default configuration: OpenAI for the language model, OpenAI for embeddings, and Qdrant storing to /tmp/qdrant. That matches Mem0's Python quickstart. It was one run on one version, and you can repeat it with pip install mem0ai==2.2.1 and python -c "from mem0 import Memory; Memory()".

What we ran on October 24, 2026: a setup with no OpenAI

Changing only the embedder was not enough, because the default language model is also OpenAI and Memory still failed at setup. We then set all three parts: fastembed with BAAI/bge-small-en-v1.5 (384 dimensions) for embeddings, a local Qdrant path for storage, and an ollama block for the language model. With no OpenAI key, Memory.from_config set up in 0.4 seconds, and add with infer=False stored the text as given and search returned it, 2.3 seconds in total including a one-time 65 MB embedding model download. The default add, which extracts memories from text with the language model, failed with an LLM error because no Ollama server was running, so a fully local Mem0 also means installing Ollama and pulling a model. Extra packages were fastembed and ollama, and the embedding size had to be stated twice. This was one run on mem0ai 2.2.1, Python 3.13, Linux.

The first-memory test, side by side

What each tool asks of one person before the first memory is saved. The Mem0 column is its default setup, from our run and its quickstart. The ModelBrain column is from our own docs, not a timed run.
Step	Mem0 open source, default setup	ModelBrain local
Runtime	Python 3.10 or newer, then `pip install mem0ai`	A desktop app for macOS, Windows or Linux. No runtime to install
Account or key	An OpenAI API key in `OPENAI_API_KEY`. In our run without one, setup failed	No account and no key for local use
Model that reads your text	OpenAI `gpt-5-mini` extracts facts by default. You can configure another provider	None for saving. An optional reasoning model for recall runs on your computer
Embeddings	OpenAI `text-embedding-3-small` by default	Computed on your computer's CPU
Where memories are stored	Qdrant files at `/tmp/qdrant` by default, plus a history database at `~/.mem0/history.db`	A vault folder on your computer
Who a memory belongs to	You pass a `user_id` on each call, as the quickstart does	One person's vault, with partitions to separate topics
Code to write	A script that calls `m.add()` and `m.search()`	None. Your assistant calls the remember and recall tools

What the count says

Mem0's default path asks you to supply or accept six things and then write code: a runtime, a key, an extraction model, an embedding model, a storage location and an identifier. ModelBrain's local path asks for none of them. You install the app and connect an assistant, which for Claude Desktop means restarting it once.

The defaults also decide where your text goes. With Mem0's defaults, the text you add reaches OpenAI twice, once for fact extraction and once for embeddings. ModelBrain makes no network call when you save, recall or delete a memory. The one exception is ChatGPT on the web, which reaches ModelBrain through a relay that handles each request in transit and does not store it. Our privacy page lists every case.

Where Mem0 is the better choice

Pick Mem0 if you are building an application and need memory for each of its users, or if you want a hosted service rather than something on your own computer. Mem0 has been in production longer than ModelBrain, which is in beta. You can also change its defaults. We did not test a Mem0 setup that runs entirely on local models, so we cannot say how many steps that takes.

Where ModelBrain costs you something

ModelBrain is built to be called by an assistant over MCP, not embedded in your own code. Its engine is not open source today, so you cannot read it the way you can read Mem0. The free local vault is not encrypted at rest, and the three readers a memory can have explains what that does and does not mean. Syncing the same memory across computers is paid, at $6 a device a month billed yearly.

For other tools in this category, see our broader comparison with Mem0 and Letta and MCP memory servers compared.

How to run the first-memory test on any tool

Start from a fresh environment with no API keys set. Follow the tool's own quickstart exactly, without borrowing anything from your current setup. Write down each thing it asks you to install, supply or choose, and the first error you hit. Stop when one memory is saved and can be read back. The count and the first error are your result.

The test has a limit. It rewards good defaults and says nothing about depth, so a tool that fails it can still be excellent once configured. What it measures is how much a person who is not a developer has to know before the tool does anything at all.

Sources

Python quickstart, Mem0 docs
Privacy: local, sync and backup, and Teams, ModelBrain
Connect ModelBrain to Claude Desktop, ModelBrain docs
Pricing, ModelBrain

Common questions

What is the best Mem0 alternative for one person? It depends on what you want. If you want Claude, Cursor or ChatGPT to remember across apps without writing code, a local MCP memory such as ModelBrain fits. If you are building an app with memory for each of its users, Mem0 is built for that.

Does Mem0 need an OpenAI key? With its defaults, yes. Memory() uses OpenAI for the language model and embeddings, and in our run on mem0ai 2.2.1 without OPENAI_API_KEY it failed at setup with a missing credentials error. You can configure other providers. In a second run on October 24, 2026, with a local embedder (fastembed), a local Qdrant path and an Ollama language model block, Memory.from_config set up in 0.4 seconds with no OpenAI key, and storing text as given worked. Extracting memories from conversation also needs a running Ollama, and without one it failed with an LLM error.

Is ModelBrain free? The local product is free with no memory cap and no account. Syncing between your computers is $6 a device a month billed yearly, or $8 monthly.

Last checked October 19, 2026. Settings and tool behavior change often, so confirm a detail on the vendor's own page before relying on it.