Some source systems allow fact data to exist without the related reference data.
For example, you may allow the creation of an invoice for a particular customer before the related record on the customer table has been created. This may seem a strange way of doing things but the situation does exist.
Not, when you are importing the invoice (in this example), you need to create the customer record in your destination database so that, for instance, foreign key constraints may be satisfied. You can only create a skeleton of the customer record based on information that is available on the invoice (which might be limited). This is the inferred member. You would need to have code that create the customer record if it doesn't exist.
At some point in the future, the customer record may exist in the source system and get imported. You are now able to add all the fields known about the customer. The inferred member support is able to deal with an existing record that was inferred from the import of another table. One technique is to include a flag on the customer table (e.g. InferredFlag) as a bit datatype. The invoice import would set this to true. When the customer import runs, it can be told able this flag and set it to false automatically.
Now, the question you may have is why do this ? It allows you to be explicitly aware of data that is incomplete yet still be able to use the data as though the record really exists.
I suggest that you search BOL or MSDN for "SSIS Inferred Member".