Customize Model Memory

Customize Model memory stores reusable facts, preferences, and corrections in a model-specific hidden knowledge base. When the Runtime receives a new question, it retrieves relevant memories and adds them to the context for that request, without fine-tuning the model.

Open the Memory page

Open a Customize Model and go to Behavior → Memory in the sidebar. Memory has five subsections:

Only the Customize Model creator, collaborators, and platform administrators can open and manage these sections. Regular members can use the model but cannot access memory management.

Enable and configure memory

In Settings, turn on Enable memory, adjust the required values, and select Save changes.

The first time memory is enabled, the system creates a hidden memory knowledge base for the model and backfills eligible existing memories into its index.

There is currently no interactive Auto-Learn switch in the UI. Instructions for enabling the old Auto-Learn setting no longer apply.

Memory sources and initial status

When memory is enabled, the Runtime may also extract explicit preference signals after a response and write them directly to the hidden knowledge base. These Runtime memories do not necessarily have a corresponding Memory entry, so they may not appear in the list or statistics.

What the Runtime uses

For content shown under Memory entries, only memories that meet all of these conditions remain in the hidden index:

• Memory is enabled for the Customize Model
• The entry is enabled
• Its status is approved
• Its content was successfully synchronized to the hidden memory knowledge base

Each request follows this general flow:

Disabling, rejecting, or deleting an entry causes the system to attempt to remove it from the index. Enabling, approving, or editing an entry causes another synchronization attempt. An index synchronization failure may return a warning without rolling back the database change.

Manage memory entries

Under Memory entries, you can:

• Add a manual memory
• Filter by status or type
• Edit memory text
• Enable or disable an individual memory
• Approve or reject a pending memory
• Delete a memory that is no longer needed
• Move through memories in pages of 20

The current search field filters only the entries loaded on the current page. It is not a server-side full-text search across every page.

Review pending memories

Under Pending review, approve or reject each pending entry. An approved and enabled entry is synchronized to the hidden index; a rejected entry is excluded from Runtime retrieval.

Approve all currently processes the pending entries loaded by the page, up to 50 at a time. If more pending entries exist, reload the section and continue reviewing them.

Process repair suggestions

When a Playground user gives a Customize Model answer a thumbs-down, the system creates a pending repair item. Repeated reports of the same pending question are consolidated when possible and increase its report count.

The system attempts to verify the question again and prepare an editable suggested answer. The suggestion may still be running or may fail. A creator or collaborator can:

1. Expand the item to review the original question and answer.
2. Review and edit the AI suggestion, or write the correct answer manually.
3. Select Submit and write to memory to create an approved memory from the verified answer.
4. Select Ignore when the item should not become memory.

Use the Feedback API

The public endpoint is POST /v1/feedback. To create a memory, the request currently needs all of the following:

• An organization or project API key accepted by this endpoint
• rating: "good"
• memorize: true
• A non-empty memory_content
• A resolvable Customize Model, identified by custom_model_id, model, or request_id

Runtime resolves the target in this order: an explicit custom_model_id / customModelId, an accessible model matching the model slug, then the customModelId recorded on the trace identified by request_id / requestId. Ordinary integrations should prefer model or the request ID returned by an earlier Runtime call.

curl -X POST https://your-runtime-base-url/v1/feedback \
  -H "Authorization: Bearer sk-your-organization-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "your-model-slug",
    "rating": "good",
    "memorize": true,
    "memory_content": "The user prefers concise bullet-point summaries.",
    "comment": "Preference verified"
  }'

A memory created through this public API is currently approved immediately and the Runtime attempts to synchronize it to the index; it does not enter Pending review first. A request with rating: "bad" returns a successful feedback response but does not create memory or enter the Playground repair queue.

Content recommendations

• Store reusable preferences, rules, corrections, or stable facts.
• Make each entry understandable without relying on missing conversation context.
• Verify content before approval and regularly remove stale or conflicting memories.
• Never store API keys, passwords, personal data, or other sensitive content that should not be sent to the model again.

To change the model's general behavior, see System prompt. For member access, see Model permissions.