As noted in issue 759958, memory cache is not really a resource cache anymore. Can you post a trace of this issue or URL that consistently repros?

As Hiroshige notes, memory cache just prunes decoded data, but I don't think script resources do any pruning.

ScriptResource is evicted from MemoryCache if it is garbage collected.
Probably, ScriptLoader::Execute() (which is used only for scripts under ScriptRunner, including async scripts) erases the last strong reference to the async ScriptResource, and thus GC can collect the ScriptResource, evicting it from MemoryCache.

Any reason why this is only happening for async scripts for FB?

I haven't tested at all, but I guess probably ScriptLoader::Execute() (used for async scripts) clears |resource_| member, while other paths for executing non-async scripts don't clear that (and |resource_| keeps alive the ScriptResource until the corresponding <script> tag is alive).

Probably, we need to keep the semantics of "Resource is alive and available on MemoryCache, as long as someone is using the Resource", and having a <script> tag explicitly is "using the Resource", so this seems to be counter-intuitive.

If the script source text is anyway referenced by V8 side, then keeping also the ScriptResource alive wouldn't cause the memory consumption regression.

I think we want to keep the semantics here as is. The issue to be addressed in different strategy as tracked in  crbug.com/766653