Skip to main content
Memory Router puts your MemoryLake memory in front of the models you already call. Keep your OpenAI or Anthropic SDK, change the base URL, and add a boundary_id to each request: relevant memories are added to the prompt before the model is called. Requests and responses keep the provider’s own format. This reference covers what Memory Router adds; for every other field, see the provider’s API reference. To get started step by step, see the Quickstart.

​
Base URL

Memory Router lives under the same base URL as the rest of the MemoryLake API, https://app.memorylake.ai/openapi/memorylake, in its own /router section.
The earlier endpoints https://app.memorylake.ai/openai/v1/… and https://app.memorylake.ai/anthropic/v1/messages keep working and behave the same. New integrations should use the paths above.

​
Endpoints

​
Authentication

Every request needs a MemoryLake API key (sk-...). Where it goes depends on whose model account pays for the call: Your provider key is forwarded to the provider and never stored. See Deployment Modes.

​
Memory

Add boundary_id as a query parameter, not a header or a body field:
A Boundary decides which memory a request uses. boundary_id applies to Chat Completions, Responses and Messages. Without boundary_id, the request goes to the model as is, with memory off; it is still authenticated and metered. See How It Works.

​
Models

Use the model’s native name, such as gpt-4o-mini or claude-haiku-4-5, with no provider prefix. List Models returns every model you can call and which base URL reaches it. The OpenAI and Anthropic /v1/models endpoints are not available.

​
Streaming

Set stream: true to receive server-sent events in the provider’s format. Events are forwarded as they arrive.

​
Tracing

Every response carries an X-Trace-ID header. Send a W3C traceparent header to make the request part of your own trace. Quote the trace id when you contact support.

​
Errors

List Models answers in the MemoryLake API format, like every other MemoryLake endpoint: errors carry success: false, a message and an error_code. See Error Codes. The inference endpoints keep the provider’s format. Errors from Memory Router itself on those endpoints: 401 and 403 use the OpenAI error shape ({"error": {"message", "type"}}). 402 returns {"success": false, "message": "insufficient quota"}. 429 returns {"success": false, "code": "AUTH_THROTTLED", "action", "error": {"type", "message"}}. 502 returns plain text. 503 returns {"success": false, "message", "code": "REQUEST_BLOCKED"}. Other errors are passed through from upstream unchanged; errors from the model provider keep the provider’s own format.