Skip to main content

​
Overview

With Bring Your Own Key (BYOK), Mem0 runs the LLM calls that turn your messages into memories on a model you choose, using your own API key. Mem0 sends your message text to that provider under your account, and your provider bills you directly. If your key stops working, Mem0 doesn’t switch to its own key. The add fails and shows why. Embeddings, search and reranking still run on Mem0.

​
Where to find it

BYOK is set per organization, in Settings → BYOK in the Mem0 dashboard. Only organization owners see this tab. One key covers every project in the organization.
Organizations created through the Vercel Marketplace are managed from Vercel and can’t use BYOK.

​
Supported providers

For Azure OpenAI, use your resource’s endpoint from the Azure portal, such as https://your-resource.openai.azure.com/.

​
Set up BYOK

1

Open the BYOK tab

Go to Settings → BYOK in the dashboard.
2

Choose a provider

Pick OpenAI, Anthropic, OpenRouter or Azure OpenAI from the Provider list.
3

Choose a model

The Model list shows the models we tested, with a Recommended tag on our pick for that provider. Under the model, the tab shows its estimated cost per 1,000 adds on your provider account.To use a model that isn’t in the list, choose Other model ID and type the model ID exactly as your provider names it. For Azure OpenAI, type your deployment name in Deployment / Model.
4

Paste your API key

Paste your key into API key. For Azure OpenAI, also fill in Endpoint URL.
5

Save the key

Click Save key. Mem0 sends a short test request to your provider with your key and model, then saves them. When the test passes, the top of the tab shows your provider and model, such as Running on OpenAI · gpt-6-luna.
Mem0 encrypts your key and never shows it again. To change only the model, pick the new model and leave API key empty to keep the saved key. To stop using BYOK, click Remove key. New adds then run on Mem0’s default model, within your plan’s limits.

​
Errors and rate limits

​
When you save a key

If the test request finds one of these problems, Mem0 doesn’t save the key: Some problems don’t stop the save, because they can be fixed on your provider’s side without changing anything in Mem0. In that case Mem0 saves the key and shows a warning at the top of the tab that names your provider. For example, with an OpenAI key:
  • No credit: “Key saved, but OpenAI says this account has no credit. Adds will fail until you add credit with OpenAI.”
  • Rate limited: “Key saved, but OpenAI is rate limiting requests right now. Adds may fail until the limit resets.”
After you add credit or the limit resets, click Check again to run the test again on the saved key.

​
When an add fails

Adds run in the background. Each add returns an event_id, and you can follow it with Get Event. When an add fails for one of the reasons below, the event ends as FAILED and its error field says why:
In the dashboard, open the failed add on the Requests page to see the same reason. Mem0 doesn’t retry these adds, so send them again after you fix the cause.

​
Rate limits and temporary errors

When your provider rate limits a request, times out or returns a server error, Mem0 retries it automatically, so the add can take longer to finish. If this happens often, raise the rate limits on your provider account.

​
Choose a model

Three things decide which model fits:
  • Accuracy: how well the memories a model writes answer questions later.
  • Cost: what your provider charges for the calls behind one add.
  • Latency: how long it takes from sending an add until its memories show up in search.
We measured these on LoCoMo, a public benchmark of long conversations, for a limited set of popular models. Your own data can give different results. Based on what we measured, these are our picks: Best value balances accuracy and price, Best under $1 is the most accurate model under $1 per 1,000 add requests, and Best quality has the highest accuracy.
The three picks. Best value: DeepSeek V4.1 Flash, open weights, on OpenRouter: 92.9% accuracy, $5.34 per 1,000 add requests, 18.5 s latency at p95. Best under $1: Amazon Nova Micro, closed, on OpenRouter: 88.0% accuracy, $0.86 per 1,000 add requests, 10.2 s latency at p95. Best quality: Claude Opus 5.5, closed, on Anthropic, OpenRouter: 93.5% accuracy, $158.55 per 1,000 add requests, 18.1 s latency at p95.
The latency above is for adds, which run asynchronously: the add call returns an event ID right away, and the memories become searchable once processing finishes. Search doesn’t use your model, and its server-side latency is about 270 ms at p50 and 640 ms at p95.
Accuracy against cost for every model we tested. The green line marks the best accuracy available at each price, the gold area holds the models under $2 per 1,000 add requests, and the gray dashed line is Mem0’s default model, at 92.5%.
Scatter chart of accuracy against cost per 1,000 add requests on a log scale, one ring per model, each labelled with its accuracy. Mistral Nemo sits at the far left at $0.68 and 84.1%; Claude Opus 5.5 at the far right at $158.55 and 93.5%. A dashed gray line marks Mem0's default model (free tier) at 92.5%. A green shaded region holds the models above it for under $10 per 1,000 add requests, a gold one the models under $2 per 1,000 add requests, and a dashed green line marks the best accuracy available at each price. Tags mark our picks: best value, DeepSeek V4.1 Flash; best under $1, Amazon Nova Micro; best quality, Claude Opus 5.5.