Memory health and confusions per store: how to read the numbers

Where to find it

In your account menu, under Order reading, Extraction memory tab. The two cards are at the top, before the Customer selector, and don't depend on choosing a customer:

• Memory health: the whole organization;
• Learned confusions by store: one store at a time, chosen in the card's own Store selector.

Both are read-only.

What the Memory health card is for

It answers two questions that would otherwise stay invisible: is Dri learning? and is what she learned helping?. The first three numbers cover the period shown next to the title ("last 30 days"); the four below are the current stock of what has been learned.

A number shown as — means there hasn't been a case in the period to calculate it yet. It isn't zero.

The three memory-effect indicators

Twin alert accuracy (switches / reviewed): when Dri flags a line with "Similar code in the catalog — confirm which one it is" and someone on the team checks it, how many times the person switched to the twin product. High means the alert is right: when it shows up, it's usually a real switch.

Alert review (reviewed / raised): of all the twin alerts raised, how many someone actually checked. Low means the team is confirming quotes without looking at the alert.

Injection reach (memory reached the prompt): of the order readings for an identified customer, in how many that customer's memory was used. Low means many customers don't have memory yet, or that readings are arriving without the customer set.

The four inventory numbers

• Active priors (e.g. "3/12"): store confusions in effect right now, out of the total recorded across all stores.
• Learned aliases: total code aliases in the memories of all customers.
• Frequent SKUs: total frequent-product entries, adding up all customers.
• Pending suggestions: suggestions waiting for Approve or Reject, adding up all customers.

If "Sample truncated (too many records) — approximate numbers." appears, the organization has too many records to add everything up, and the numbers are an estimate.

What a confusion learned per store is

Each time someone on the team corrects a line (switches the product by hand or resolves the open question by choosing another product), the store records the pair "the document said X, the right product was Y". This applies to all of the store's customers, including a new customer with no memory at all.

An active confusion doesn't switch the product on its own. It only makes Dri raise the "Similar code in the catalog — confirm which one it is" alert when she reads that code and a twin product exists in the catalog.

How to read a store's confusion table

Choose the store in Store. Each row is a pair, from most influential to least:

The note at the bottom of the card states the cutoff: a confusion becomes active when the decayed weight is at least 2. In other words, a single correction never activates a confusion: the same correction is needed at least twice, recently.

Why a confusion becomes inactive on its own

The weight halves every 90 days without the same correction repeating. A pair corrected three times at the start of the year and never again drops out of the alert over time, without anyone having to delete anything. This is on purpose: if the catalog was recoded or the customer stopped making the mistake, the old confusion shouldn't keep firing.

There's no button to delete a confusion. If a pair is firing an undue alert, the review itself fixes it: confirming the product read in the alert, instead of switching, adds no weight to the pair.

Common questions

The store has no learned confusions. The screen says "This store has no learned confusions yet.": no one has corrected one of its lines by switching the product yet.

Why does the confusion show n = 5 and is inactive? The five corrections are old; the decayed weight dropped below 2.

I corrected once and the alert didn't show on the next order. That's right: a single correction adds weight 1, below the cutoff of 2.