To a sufficient degree of simplification, the transistor obeys the Ebers-Moll equation:
Ic = Is * exp(Vbe/Vth)
Where Is is the saturation current (roughly proportional to the off-state or reverse leakage current: for 2N3904 and the like, usually in the pA range; proportional to device size, temperature dependent), Vbe is the applied base-emitter voltage (or emitter-base for a PNP -- PNP works exactly identical to NPN, just flip all the arrows), and Vth is the thermal voltage (KbT/q_e ~= 26mV, give or take a factor (emission coefficient N)).
But discrete transistors have wildly varying parameters: even if you match Vbe (by wiring the bases and emitters together), the thermal voltages may vary (the "output" side of the mirror may have much higher voltage drop, dissipating much more power than the "input" side, which is dropping only Vbe -- as a result, the output side heats up), and the chips themselves may not be matched (in geometry, doping and so on), resulting in different Is values.
By using emitter resistors, you potentially simplify the circuit much more. Instead of requiring Ebers-Moll, you can describe the circuit fairly well (meaning, within a certain accuracy over a certain range of operation) with the even more basic linear-offset model (Ic = Ib * hFE, Vbe = either 0.7V, or v_be + Ib * r_b, using the incremental resistance). By not having to invoke Ebers-Moll for basic operation, you can immediately guess that the process and temperature variation will be significantly better, and this is exactly the effect.
By the way, suppose you only added the emitter resistor to one side. Now you get a hybrid. One side looks more-or-less linear, but the other side is still very nonlinear. What's the result? A logarithm (or exponent, depending on which side gets it). This is the Widlar Current Mirror, which has some applications as a crude bias reference (the output is still dependent on the input, but it varies logarithmically as Vbe vs. Ib, which is better than a resistor which varies proportionally) and computation element (the exponent of the sum of two logs is the product of those numbers, i.e., you can build an analog multiplier of sorts, making Q = X*Y by taking Q = exp(log(x) + log(y)), and of course, currents are easy to add by summing them in a single node).
Tim