What 13 days of phone notifications look like
My phone has been running the memento collector for about two weeks. The sample is small but real: 1873 notifications over 13 days, every one encrypted in transit and at rest. This note records what the stream actually looks like, because anyone building on notification streams should know these shapes before designing.
Concentration
| app | count | share |
|---|---|---|
WeChat (com.tencent.mm) | 1766 | 94.3% |
Alipay (com.eg.android.AlipayGphone) | 107 | 5.7% |
Collection was enabled for exactly these two apps, so the concentration is a property of the collection policy, not of the phone. That is itself the first finding: a notification stream is a sample, not a census — any analysis of it is also an analysis of what you chose to capture.
Within the two apps, the two streams mean different things: the WeChat share is a social graph, the Alipay share is a transaction log. Two apps already give a usable skeleton of a day.
Daily variance
Daily counts, in order:
23 77 255 77 218 155 331 155 122 331 39 201 39 167Mean 144/day, median 155, range 23–331 — a 14x swing between the quietest and loudest days. The spikes are not noise: they are single group conversations going active, or batches of payment confirmations. For a memory pipeline this matters, because it means the interesting information is in the variance, not the average — a system sized for the mean will be idle most days and saturated exactly when something worth remembering happens.
Duplicates
Early tests measured ~38% of the stream as repetition: the same message re-posted, apps re-sending the same alert, status updates for one underlying event. Any consumer of this data needs dedup before anything else — embedding and summarizing duplicates directly pollutes whatever memory store sits downstream.
Caveats
n=1 device, 2 apps, 13 days. This is a characterization of one stream, not a population study. The transferable findings are structural: collection policy defines the sample, daily variance spans an order of magnitude, and roughly a third of the volume is duplicate. Design for those three shapes and the specifics stop mattering.