Noun phrases involving relative clauses are assumed to universally comprise syntactic “islands” for extraction, but Swedish exists as a possible exception. Using eyetracking while reading, we investigated whether extraction from Swedish restrictive relative clauses (RCE) ([such old wheelbarrows]1 saw I a man that always washed __1 with benzine...) elicit similar processing costs as extractions from non-restrictive relative clauses, which are known to comprise strong islands (StrongIs); or if they pattern closer to extractions from non-island constructions (NonIs). We also examined to what extent non-linguistic variables (working memory WM, verb-object frequency, and pragmatic-fit) contribute to such differences. Results from a mixed models analysis of the embedded verb (washed) and spillover region (with...) suggest that in early measures, both RCE and NonIs show facilitation relative to StrongIs, but in late measures, RCE patterns closer to StrongIs as WM and pragmatic-fit increase, suggesting that Swedish RCE acceptability is partly dependent on non-linguistic factors.