All Products
Search

This Product

    PolarDB:PolarDB Agent Memory billing

    Document Center

    PolarDB:PolarDB Agent Memory billing

    Last Updated:Sep 17, 2026

    PolarDB Agent Memory uses the pay-as-you-go (postpaid) billing method. This topic describes the billable items, unit prices, and billing cycles of PolarDB Agent Memory. Use this information to estimate your costs before you create PolarDB Agent Memory and to control costs by cleaning up memory libraries that you no longer use.

    Billing method

    Pay-as-you-go (postpaid) means you use the service first and settle the fees after the bill is generated. The billing cycle is hourly. The system calculates the actual usage of all billable items on an hourly basis and generates a bill. After you create PolarDB Agent Memory, fees are generated every hour. Fees continue to accrue for as long as the memory library is retained. If you no longer need a memory library, clean it up in a timely manner. The following tables describe the billing dimensions, unit price units, and region-dependent pricing of the two billable item categories.

    Billable item

    Billing basis

    Unit price unit

    Region-dependent

    Compute and storage resource fees

    Number of memories and duration

    /10,000 memories/hour

    Yes

    Model fees

    Tokens consumed by model invocations

    Per million tokens

    Yes

    Compute and storage resource fees

    After memory extraction, memories are vectorized and streamed into the index, which continuously occupies compute and storage resources:

    • Compute resources: the compute resources used to create, modify, retrieve, and delete memories and to build memory indexes.

    • Storage resources: the storage required to vectorize extracted memories and to store vectors and vector-related data.

    • The size of compute and storage resources is proportional to the number of memories. The system automatically scales in and out based on the actual usage.

    Standard Edition

    Standard Edition is billed based on the number of memories and duration. It supports 30 QPS and up to 20 million memories/vectors.

    Region

    Unit price

    The Chinese mainland

    0.00217 USD/10,000 memories/hour

    China (Hong Kong) and overseas

    0.00369 USD/10,000 memories/hour

    Model fees

    Fees for memory extraction, memory vectorization, and text reranking model inference are calculated based on the tokens consumed by the actual model invocations. PolarDB Agent Memory supports the following models:

    Note

    PolarDB Agent Memory uses models on Alibaba Cloud Model Studio (for example, qwen3.6-plus, text-embedding-v4, or qwen3-rerank).

    • Text generation: qwen3.6-plus, glm-5.1, kimi-k2.5, and kimi-k2.6.

    • Text embedding: text-embedding-v4 and text-embedding-v3.

    • Text reranking model: qwen3-rerank and gte-rerank-v2.

    The following tabs describe the detailed billing rules of each model category.

    Text generation

    When you use the memory library and invoke the large language model service to extract memories, fees are calculated based on the input tokens and output tokens.

    Qwen-Plus

    Model name

    Region

    Input token range per request

    Input unit price (USD/per million tokens)

    Output unit price (USD/per million tokens)

    Non-thinking mode

    Thinking mode (chain-of-thought + answer)

    qwen3.6-plus

    The Chinese mainland

    0<Token≤256K

    0.3312

    Explicit input cache: 0.4128

    1.9812

    1.9812

    256K<Token≤1M

    1.3212

    Explicit input cache: 1.6512

    7.9224

    7.9224

    China (Hong Kong) and overseas

    0<Token≤256K

    0.3312

    Explicit input cache: 0.4128

    1.9812

    1.9812

    256K<Token≤1M

    1.3212

    Explicit input cache: 1.6512

    7.9224

    7.9224

    GLM

    Model name

    Region

    Mode

    Input token range per request

    Input unit price (USD/per million tokens)

    Output unit price (USD/per million tokens)

    glm-5.1

    The Chinese mainland

    Non-thinking mode and thinking mode

    0<Token≤32K

    0.99

    • Input cache: 0.198

    • Explicit input cache: 1.2372

    3.9612

    32K<Token≤200K

    1.32

    • Input cache: 0.264

    • Explicit input cache: 1.65

    4.6212

    Kimi

    Model name

    Region

    Input token range per request

    Input unit price (USD/per million tokens)

    Output unit price (USD/per million tokens)

    kimi-k2.5

    The Chinese mainland

    -

    0.6888

    • Input cache: 0.138

    • Explicit input cache: 0.8616

    3.6132

    China (Hong Kong) and overseas

    -

    0.6888

    • Input cache: 0.138

    • Explicit input cache: 0.8616

    3.6132

    kimi-k2.6

    The Chinese mainland

    -

    1.1288

    • Input cache: 0.2146

    • Explicit input cache: 1.3409

    4.6889

    Text embedding

    Text embedding models convert text data into vector data for memory retrieval. Fees are calculated based on input tokens; output is not billed.

    Model name

    Region

    Input unit price (USD/per million tokens)

    text-embedding-v4

    The Chinese mainland

    0.0868

    China (Hong Kong) and overseas

    0.0893

    text-embedding-v3

    The Chinese mainland

    0.0868

    China (Hong Kong) and overseas

    0.0893

    Text reranking model

    Text reranking models provide relevance ranking. Fees are calculated based on input tokens; output is not billed.

    Model name

    Region

    Input unit price (USD/per million tokens)

    qwen3-rerank

    The Chinese mainland

    0.0868

    China (Hong Kong) and overseas

    0.0868

    gte-rerank-v2

    The Chinese mainland

    0.1389

    China (Hong Kong) and overseas

    0.1389

    References