How to choose the model that writes your memories

Every add request to Mem0 runs a language model. It reads the conversation and decides what to keep: a user’s preferences, a deadline, a change of plans. With bring your own key (BYOK), you can pick that model.

The model choice impacts three numbers:

We ran 17 models through the same benchmark to see how each performs with the Mem0 algorithm, and naturally, the spread is quite wide. The least accurate model misses 2.4 times as many questions as the most accurate. The most expensive costs 233 times as much as the cheapest.

This article covers what BYOK gives you, which providers it works with, and how to decide whether a model fits your use case.

What BYOK gives you

BYOK works with four providers: OpenAI, Anthropic, Azure OpenAI and OpenRouter. OpenRouter carries every model we tested, including open-weight models such as DeepSeek, GLM, Gemma and Mistral, and the Google and Amazon models.

What we measured

We ran every model on LoCoMo, a public benchmark of 10 long conversations and 1,540 questions about them, and recorded three numbers:

Accuracy ranged from 84.1% to 93.5%. Cost ranged from $0.68 to $158.55 per 1,000 add requests. Latency ranged from 6.1 to 86.5 seconds. Mem0’s default setup scores 92.5%.

Five questions to pick a model

1. How much does a missed memory cost you?

Start with what a wrong or missing memory does to your product. If your app tracks health details, long-running goals or account history, one missed fact can lose a user’s trust. If memories only make replies a little more personal, a miss costs less.

At 93.5%, the best model misses about 1 question in 15. At 84.1%, the weakest misses about 1 in 6. Six models score above Mem0’s default: Claude Opus 5.5, Claude Sonnet 5.5, GLM-5.3 Flash, DeepSeek V4.1 Flash, GPT-6 Luna and GPT-5 mini.

The top of the range is flat. Claude Opus 5.5 scores only 0.6 points above DeepSeek V4.1 Flash and costs about 30 times as much. Claude Sonnet 5.5 scores within 0.1 points of Opus at half the price. If accuracy comes first, the last point costs much more than the ones before it.

2. What does it cost at your volume?

Multiply the price per 1,000 add requests by your monthly volume. At 100,000 add requests a month:

At high volume, the cheap end matters. Amazon Nova Micro is the most accurate model under $1 per 1,000 add requests among our test set. Amazon Nova Lite adds 1.3 points for about twice the price.

3. How soon does a new memory need to be findable?

Mem0 extraction runs asynchronously so that a slow model won't delay your app’s reply. But a new memory can’t be found until it’s written.

If your users refer back to something they said a few messages earlier, choose a fast model. Among the models above Mem0’s default setup, Claude Sonnet 5.5 (10.4 s) and GPT-6 Luna (15.2 s) are the fastest. Models that reason before they write are slower: Gemini 3.8 Flash takes 36.0 s, GPT-5 mini 37.7 s and GLM-5.3 Flash 86.5 s.

If your app only reads memories in the next session, latency doesn't matter as much.

4. Which provider are you already on?

5. Does it hold up on your own conversations?

LoCoMo is long, casual chat between two people. Your traffic may look different: short support tickets, health check-ins or several languages. Treat the benchmark as a suggestive shortlist, then test the shortlist on your own data.

Our picks

Setting it up

An organization owner can go to Settings > BYOK, pick the provider and the model, and paste the API key. From then on, Mem0 extracts memories with that model and bills the calls to your key. Your Mem0 integration code stays the same.

If you run your own comparison and get different numbers on your data, we want to hear about it.

Written by: @Rudrajmehta, Product, Mem0