Well, yes, that comes back to the online guiding topic a bit - for example, the “Practical Path Guiding” paper caches a spatial radiance estimate which could be used for secondary bounces (they use it for a heuristic that determines path splitting/termination).
Something like that could be added to Cycles, but there are a few implementation questions - for example, the cache is built during rendering, so we’d need to either do a prepass or do some hybrid scheme where the GPU renders as usual while the CPU processes its output and updates the cache. Even then, the noise from the first few samples where no cache is available yet would remain in the image.
Also, as soon as you start to use the cache to directly look up radiance instead of sampling distributions, issues like light bleeding start showing up again. Sure, you can address those, but at that point it starts to become really painful to implement all of that.
Personally, I’d rather stick with denoising approaches and e.g. apply stronger filtering to the indirect light transport components. CNN-based denoising algorithms show great results when it comes to spreading out noisy low-frequency images without blotchy artifacts, so doing e.g. a hybrid scheme with a NFOR-based direct lighting filter and a CNN-based indirect lighting filter might make sense.