The cold-start problem only affects CF approaches.

Memory-based CF operates on the full user-item-rating matrix to create recommendations.

In case of binary ratings (i.e., user interacted with item, or did not), no user bias factor is required when computing memory-based CF.

Pure Collaborative Filtering approaches do not require any domain-specific knowledge.

Memory-based CF tries to explain user ratings by so-called latent factors in a low-dimensional space.

Item-based CF (in the memory-based variant) tends to scale better for several reasons, including: (1) item-item similarities can be computed offline and updated from time to time and (2) 

only

items the active user rated are considered when identifying nearest neighbors.

Comparing memory-based with model-based CF, the recommendations created by the former are generally easier to interpret.