Agree.
That is, the narrower your receiver, the less noise it picks and creates. I built a magnetic receiver at 457kHz that had a noise bandwidth of about 1Hz (triple heterodyne), its sensitivity was about -170dBm: enough to pick the 10.000th harmonics of the mains. It worked like an ARVA (no idea how you call it in English: it locates the magnetic transmitter hold by a victim entrapped in an avalanche) but mine had a range of 100m.
The basic reason is that the physical fluctuation is an energy (kT). If you observe this fluctuation over a broader bandwidth, that is a shorter time, the corresponding power is bigger.
Now, it is difficult to now precisely what the noise bandwidth is. It is not the -3dB bandwidth, but the integral of the power response of whatever filters the frequencies. For instance, some spectrum analyser's doc give the noise bandwidths of the IF filters; other analysers have software to convert to dBm/Hz (should be : per log(Hz)...); bad ones have nothing.
If you think of a spread spectrum receiver, you first have a broad bandwidth with much noise, and after pulse compression (call it despreading if your job wants it) a narrow bandwidth with hopefully less noise. Well, sometimes. Because then, the signal-to-noise ratio before pulse compression would generally be negative, and few receivers still work under these conditions - which means that spread spectrum often has a poor sensitivity.
Several effects are less obvious. For instance in a heterodyne (not a Q-I one), noise at the image frequency, picked at the antenna or created by the preamplifier is filtered out at the input or output of the preamplifier, but noise added by the mixer is not. So mixers have a bad noise figure.
You may understand an optimum demodulator as a means of reducing the noise bandwidth.
Also think of radio-astronomy correlation receivers. They use two or more receivers and preferable antennas. Instead of adding the signals, they integrate their product over a (very) long time but keep the input bandwidth large. This filters out very efficiently the uncorrelated noises produced by the receivers. The equivalent noise bandwidth is much smaller than the input bandwith, but is not the inverse of the integration time - one wins like sqrt (F*T). This gets complicated because the process isn't linear.
Nothing obvious here. Noise is subtle anyway.