Long-term memory

Create in the console

1. Create a long-term memory store.

   When you enable the long-term memory feature for an agent application in Alibaba Cloud Model Studio Applications (see ① in the figure below), a long-term memory store is automatically created. The current application is bound to this memory store by default and uses it for memory retrieval and storage. To bind a different long-term memory store, see Call a long-term memory store in an application.

2. Obtain the Memory ID of the bound memory store.

   On the right side of the page, click the icon (see ② in the figure above). Then, click Memory ID (see ③ in the figure below) and record the Memory ID of the bound memory store (see ④ in the figure below) for future calls.

3. Memory segment information is stored and isolated by a Memory ID. To achieve more granular memory isolation, flexibly manage different long-term memory stores, and prevent memory confusion between different scenarios or users, click Create Memory ID (see ⑤ in the figure above) to create a new ID. Pass this ID when you call the long-term memory store in the application.

Create using an API

Method 1: Call the CreateMemory API operation to create a new long-term memory store. Record the memoryId (the Memory ID) from the returned parameters for future calls.

Method 2: Call the ListMemories API operation to retrieve a list of existing long-term memory stores. Record the memoryId (the Memory ID) from the returned parameters for future calls.

Create in the console

1. Click the icon next to the Long-term Memory switch. In the Memory Variable panel on the right, click Add Field to configure the fields.

   

   Parameter	Description		Configuration example
   Field name	The name of the memory variable field. Enter a custom name. The field name should be semantic to ensure that users can quickly understand its meaning and that machines can easily perform label extraction and aggregation and analysis.
   Field description	A natural language description of the memory variable field. This description prompts the large model about the field's purpose. Enter a custom description.
   Use model inference	Specifies whether to automatically update the field's value using model inference. The switch status does not affect the content extracted into memory segments.	Switch is on : The value of this memory variable is extracted from memory segments. As user information is updated in the conversation, this value is also automatically updated. This setting is recommended for subjective information, such as user preferences. Examples include dietary preferences and exercise habits. This information is not static and may change over time. The system should be able to update it dynamically. Ensure the stored information is current by periodically asking questions or analyzing user behavior.
   		Switch is off : The value of this memory variable is not automatically updated. You can enter the value manually or pass it through an API. This setting is recommended for objective information, such as user features. Examples include gender and name. This information is stable and does not change frequently. Periodically review and clean this data to ensure its accuracy and timeliness. Update or delete irrelevant or outdated information to avoid misleading the system's decision-making process.

2. Click OK to complete the configuration. You can view the configured memory variable fields in the Long-term memory area.

   

3. To publish the application, click Publish.

   You must publish the application after adding a memory variable field for the change to take effect.

4. Enter the following user queries to start a conversation. After about 1 minute, click the icon on the right side of the page to view the memory segments and memory variables that were automatically extracted from the user queries.

   User query	Content in memory segment		Value of memory variable
   I like seafood.	User likes seafood.		When the Use model inference switch for the "Dietary preference" field is turned on, the system automatically extracts the value of the memory variable from memory segments, such as "seafood, pasta, does not eat rice, does not eat spicy food, especially Sichuan cuisine".
   I love pasta and don't eat rice.	User loves pasta and does not eat rice.
   Sichuan cuisine is too spicy for me.	User does not like spicy food, especially Sichuan cuisine.
   I play badminton every weekend.	User plays badminton every weekend.		When the Use model inference switch for the "Exercise preference" and "Age" fields is turned off, the system cannot automatically extract their values, leaving the fields empty. You can manually enter a value for "Exercise preference", such as "swimming". In subsequent conversations, the memory segments and memory variables are recalled, and the model generates a response based on its judgment. For an example, see Modify the value of "Exercise preference".
   I am 20 years old.	User was born in 2005. (Inference time: 2005)
   It's raining heavily today.	Not personalized user data. No content is extracted.

5. When the user enters a new query, the system answers the question based on the memory variables combined with the prompt.

   

6. Manually enter a value for "Exercise preference", then clear the history and start a new conversation.

   

   If no value is entered for "Exercise preference", the large model bases its answer on the content of the memory segment.	If you manually enter "swimming" for "Exercise preference", the memory segments and memory variables are recalled in subsequent conversations. The model then uses this information to generate a response.

Create using an API

Method 1: Call the CreateMemoryNode API operation. In the content parameter, enter the following examples to create memory segments. If the provided Memory ID is already bound to an application in the console, you can view the content in the memory segments in the console after creation. Otherwise, you can only view the content by calling the GetMemoryNode or ListMemoryNodes API operations.

Examples:

• User likes seafood.

• User likes pasta and does not eat rice.

• User cannot eat food that is too spicy.

• User plays badminton every weekend.

• User is 20 years old.

Method 2: When using the DashScope API Reference, pass the Memory ID and enter the following examples in the prompt. The system automatically creates memory segments in the specified long-term memory store based on the prompt. In the DashScope API Reference, you can search for "long-term memory" to view code samples and detailed calling methods. After creation, you can view the content in the memory segments in the console.

Examples:

• I like seafood.

• I love pasta and don't eat rice.

• Sichuan cuisine is too spicy for me.

• I play badminton every weekend.

• I am 20 years old.

Call in the console

1. In Alibaba Cloud Model Studio Applications, find your agent application and click Configure.

2. When you enable the long-term memory feature, a long-term memory store is bound to the current application by default. Therefore, you do not need to rebind it for retrieval and storage. Simply start a session. To bind a different long-term memory store, click the icon, click Memory ID, select a Memory ID from the list, and click OK.

   A Memory ID that is already in use by another application cannot be deleted.

3. Publish the application.

4. When you start a new session, the system automatically recalls the relevant memory content based on the Memory ID bound to the application. It then passes this content, along with the current user query, to the model to generate an answer.

Call using an API

1. Obtain the Memory ID.

   For instructions, see Step 1: Create a long-term memory store and obtain the Memory ID.

2. In the DashScope API Reference, search for the "long-term memory" parameter to view code samples and detailed calling methods.

   When you call the agent application, the system automatically recalls the relevant memory content based on the provided Memory ID. It then passes this content, along with the current user query, to the model to generate an answer.

   Because the long-term memory store is bound to the application by default when the feature is enabled, if you do not pass a Memory ID, the system continues to use this default memory store for retrieval and storage.