Today we’re launching bring your own key (BYOK) on the Mem0 Platform. You choose the model that extracts your memories, and its usage is billed to your own API key.

That model reads each conversation and decides what to remember, so your choice affects three things: how accurate your memories are, what each add request costs and how soon a new memory can be found.

To help you choose, we evaluated well-known open-weight and closed models that work with BYOK, from frontier models to options under $1 per 1,000 add requests. They differ widely. The least accurate misses 2.4 times as many questions as the most accurate, and the most expensive costs 233 times as much as the cheapest. Below are our recommendations, then the results behind them.

Useful links:

BYOK setup: https://docs.mem0.ai/platform/features/byok
Add your key: https://app.mem0.ai/dashboard/settings?tab=org-llm

Choosing a model

The right model depends on what matters most for your use case. These are our picks:

Best value: DeepSeek V4.1 Flash. If you’re open to any provider, start here. It scores 92.9%, above Mem0’s default model, for $5.34 per 1,000 add requests, which keeps high-volume use cases such as companions and coaches affordable. GPT-6 Luna costs less, $3.55, but scores 92.6%; GLM-5.3 Flash scores 93.0% but is the slowest model we tested, 86.5 s at p95.

Best under $1: Amazon Nova Micro. For use cases where cost comes first. It is the most accurate model under $1 per 1,000 add requests, at $0.86, and scores 88.0%, below Mem0’s default.

Best quality: Claude Opus 5.5. For use cases where accuracy comes first. It is the most accurate model we tested, at 93.5% for $158.55 per 1,000 add requests. Claude Sonnet 5.5 scores 93.4% for half the price.

Latency and the provider you already use can also change your choice.

Latency. Memory extraction runs asynchronously, so even a slow model doesn’t hold up your app, but a new memory can’t be found until it’s written. If your users often refer back to something they said a few messages earlier, choose a fast model. Among the models above Mem0’s default, Claude Sonnet 5.5 (10.4 s at p95) and GPT-6 Luna (15.2 s) are the fastest.

Provider. Each provider that BYOK supports has a model above Mem0’s default. On OpenAI or Azure OpenAI, choose GPT-6 Luna. It matches GPT-5 mini’s 92.6% at a fifth of the cost, $3.55 vs $17.82 per 1,000 add requests, and extracts memories more than twice as fast. On Anthropic, choose Claude Sonnet 5.5 or Opus 5.5 if accuracy comes first and Claude Haiku 5.5 if cost does; Haiku scores 92.3%, just under Mem0’s default, for $4.93. OpenRouter carries every model we tested.

Results

GPT-6 Luna, DeepSeek V4.1 Flash and GLM-5.3 Flash score above Mem0’s default model for under $10 per 1,000 add requests (the green area). Claude Opus 5.5 and Sonnet 5.5 are the most accurate overall. Under $2 (the gold area), Amazon Nova Lite is the most accurate, at 89.3% for $1.82, 1.3 points above Nova Micro at about twice the price. The gray dashed line is Mem0’s default model, at 92.5%, and the green line marks the highest accuracy available at each price.

What it costs a month

At 10,000 add requests a month, Amazon Nova Micro costs $8.60, DeepSeek V4.1 Flash $53, Claude Sonnet 5.5 $784 and Claude Opus 5.5 $1,586. Set your monthly volume and budget below to see the most accurate model you can afford.

How fast each model saves a new memory

Most models make a new memory searchable within 20 seconds at p95. GLM-5.3 Flash, GPT-5 mini and Gemini 3.8 Flash take longer because they reason before they write. The fastest, Gemini 3.5 Flash-Lite, takes 6.1 s but scores 89.9%, below Mem0’s default.

Every model we tested

Accuracy, cost and latency for every model, sorted by accuracy.

How we tested

We tested every model on LoCoMo, a public benchmark of 10 long conversations and 1,540 questions about them. Only the model that extracts memories changed; everything else stayed the same.

Accuracy is how often questions about past conversations are answered correctly from the memories the model wrote, graded against LoCoMo’s reference answers. Each score is the average of two runs.

Cost is what your provider bills for 1,000 add requests, estimated from each model’s calls. Each add request in this test carried about 8 messages; larger requests cost more.

Latency is how long memory extraction takes: the time from an add request until its memories can be found. We report p95: 19 of every 20 requests finish within that time. It also depends on your provider and plan.

Get started

To switch, an organization owner opens Settings → BYOK, picks the provider and the model, and pastes the API key. Mem0 extracts your memories with that model from then on, billed to your key, and your add and search code stays the same.

Add your key in the Mem0 dashboard to get started. The setup guide has the steps for each provider.