@ Green: I answered your question in this topic already:
http://warlight.net/Forum/Thread?ThreadID=35330Copy-paste:
@ Green: there's no closed-form solution for this, but basically every game changes your rating when it finishes (the mean goes down when you lose, up when you win, the standard deviation goes mostly down unless the result was very unexpected). At that moment, you can only give the influence of the most recent games, the influence of all previous games has already been lowered and cannot be traced back anymore.
Your new rating (mean) is used for the next game. Since the only thing that counts in the future, is the new mean, it does not matter anymore how you got that mean.
Since the mean is always adapted by the latest game, the influence of games is technically exponentially decreasing, this implies that in theory any more recent game has more influence than any less recent game. (in practice, a win against an opponent who is much stronger will always have a bigger effect than a win against an opponent who was weaker.) It is possible to simulate this actually, Fizzer has developed a tool for this:
http://blog.warlight.net/index.php/2012/01/trueskill/Based on my simulations with TrueSkill, the most recent games really make the difference. So depending on why you call "neglible", this actually happens pretty quickly.
Example: if you lose against a weak player, your mean (and rating) drop quite steeply. Winning against 2 players of your previous mean will be enough to put your rating back where it was. (Whereas in BayesElo those 2 players and you would each still have a slightly lower rating.)