I'm concerned that this may lead to OOM errors and/or memory leak-like behavior.

I'd say unit-tests typically don't send enormous payloads, or make thousands of requests.

Anyhow: remembering 2000 get-requests getting "foo" as answer: https://gist.github.com/sezanzeb/d5b89e10bdebf2cfb6192983743aa9eb shows 101MiB of memory usage in my system monitor. Without any requests 48MiB are needed.

Also, requests should be cleaned up after every test, if the interceptor isn't created in beforeEach callbacks. I just realized this still needs to be implemented, like https://jestjs.io/docs/mock-function-api#mockfnmockclear

I agree with you, but users may use Nock in unexpected ways. WDYT about saving only the last ten requests?

If there is a warning being written to the console when the maximum number of requests is stored, and if it is configurable, it would be an acceptable compromise I think.

Sounds good.

Alternative:

nock('http://example.test')
  .persist()
  .get(/\/dsad2.*/)
  .remember(10)  // the argument is null by default: remembers all requests

without .remember(), no requests are saved.

I think this is nice and fits well into nock, and is optional, so existing tests won't remember requests. Therefore the limit can be infinite by default. What do you think?

Adding features is easy. Maintaining them is something else :)
I know that using persist is a known pattern with nock, but can you please share with me why you use it instead of regular interceptors?
Don't get me wrong—I'm in favor of inserting this feature to nock, but I think we need to be careful how we implement it.

You can also do

  it('remembers a request', async () => {
    const interceptor = nock('http://example.test')
        .get('/foo')
        .remember()

    interceptor.reply(202, '1')
    await sendRequest('example.test', '/foo')
    expect(interceptor.requests).to.have.lengthOf(1);

    interceptor.reply(202, '1')
    await sendRequest('example.test', '/foo')
    expect(interceptor.requests).to.have.lengthOf(2);
  })

without persist.

I have commited that new test

I guess the request object should be remembered in the scope as well:

    const scope = nock('http://example.test')
        .get('/foo')
        .remember()
        .reply(202, '1')

    await sendRequest('example.test', '/foo')

    console.log(scope.request)

Though I'm somewhat confused with scopes and interceptors. A scope has a list of this.interceptors. But doing a single .reply(202, '1') on an interceptor creates a scope for a single request, which suggests that a scope belongs to a single interceptor.

ClientRequest does not have a friendly API (think about getting the body). Maybe we should transform it to a Request instance?

Do you know how?