- Change the base URL from
https://openrouter.ai/api/v1tohttps://nemonist.com/api/v1. - Use a nemonist key (
sk-nm-...) instead of your OpenRouter key.
provider, models fallbacks) work as before.
Turn memory off for an app
Memory is for coding agents. An app that calls the API (a SaaS backend, a classifier, a function-calling bot) should use a key with memory turned off. Requests under that key are sent as your code wrote them, with no memory and no added tokens, even when they carrytools, so prompt tokens and cost match OpenRouter’s.
- New key: on the API keys page, choose New key and check No memory, for an app that calls the API.
- Existing key: open the key, and under Memory choose Turn off. Turn it back on the same way.
- From code: send
"memory_off": truetoPOST /keysorPATCH /keys/{hash}with a management key.
tools is treated as a coding agent’s and is asked to link a project, which adds tokens to its prompt.
Keep paying OpenRouter directly
To keep billing to your OpenRouter account, add your OpenRouter key on the workspace’s Your key page. Requests are then billed to that account. See Bring your own key.What is different
- Memory. Requests from coding agents get their project’s memory and learn from every turn; a request with
toolscounts as a coding agent’s. Other requests, and every request under a key with memory off, are answered as OpenRouter would answer them. See Memory. - Error bodies. Status codes are OpenRouter’s, and the body keeps
error.codeanderror.message. nemonist does not return OpenRouter’s top-leveluser_idor itserror.metadatadetails (provider_name,raw,provider_error_code,is_byok,routing_funnel,failed_routing_step,ineligibility_reasons,limit_source,remedy_hint). When it knows the kind of error,error.metadataholds onlyerror_type. The message says what went wrong in words. See Errors. - Allowed models. A request for a model the key’s allowed models exclude gets
403(“This key may not use the model …”), where OpenRouter answers404. A guardrail refusal is also403, witherror.metadatanaming the guardrail. - Latency and compression. Each request passes through nemonist, which adds about 0.15 to 0.3 seconds, before the first token too. Occasionally a provider takes 2 to 6 seconds longer to answer a short request. This happens through OpenRouter directly at the same rate (about 1 to 4 in 100 calls in our tests, up to 5.6 seconds), and nemonist does not add it. If a call sits on a path with a tight timeout, such as a classifier on a user’s request, set the timeout to 10 seconds, or cut it off at 6 seconds and retry once. Replies are not compressed (no
content-encoding: gzip). - Cursor. Cursor uses its own endpoint,
https://nemonist.com/api/v1/cursor. See Cursor. - Memory endpoints.
GET /memory-fileandDELETE /memoryread and erase memory.