What is Church Memory, and why does week ten sound better than week one?

A glossary is what a church tells the system. A memory is what the system notices, keeps, and shows the church for approval: the names it says every week, the ministries it runs, the phrases it repeats. The first is right from day one; the second is why week ten sounds more like the church than week one.

The problem every Sunday starts with

Machine translation is trained on the world's text, not on your church. It has never heard your pastor's name, does not know that “the Hub” is a building and not a metaphor, and will translate a phrase your congregation has said the same way for twenty years slightly differently every week. None of this is a mistake in the usual sense. It is the absence of a memory.

A human interpreter solves this by being there every Sunday. A briefing sheet solves it for the substitute. Software has to solve it another way, and there are two.

A glossary, and what it is good for

A glossary is a list a person writes: this name is spelled this way, this term is translated with that word, this phrase stays as it is. It is exact, it is authoritative, and it is right the moment it is typed. Its weakness is that somebody has to type it, and to think of everything in advance, and the people who could do that are the people already carrying the sound desk and the children's work.

For anything that reaches the live voice, a glossary is the right instrument: a substitution the interpreter performs on speech, in front of the congregation, should be one a human approved. Ask a vendor whether their glossary is versioned, who approves a version, and whether a live service can read one that nobody has approved yet.

A memory, and why it has to be different

A memory is what the system learns on its own after each service: this name came up again, this ministry was mentioned in three services, this phrase recurs. Unlike a glossary it must be allowed to gather evidence without a human in the loop, or it is just a glossary with extra steps. That makes it a different kind of thing, and it needs different rules.

One sighting is noise. A visiting speaker's name, a one-off illustration: stored, shown as “not used yet”, and kept out of every translation until it recurs.
Recurrence is confidence. A thing observed in several separate services is likely true of the church; a thing observed once is not.
What the church typed always outranks what the system learned, and learning can never change it.
Anything learned can be pinned, switched off, or corrected — and a correction turns it into something the church said.
Erase a service, and the evidence it contributed goes with it.

The last rule matters under the GDPR: a memory built from erased services is a memory of data that no longer exists. Ask how a vendor's memory behaves when a recording is deleted.

What both of them feed

The obvious beneficiary is the live translation: the name pronounced right, the ministry not translated literally, the recurring phrase rendered the way the church renders it. The less obvious one is everything written afterwards. A summary, a manuscript, a newsletter draft written with the church's own vocabulary reads as the church's; written without it, it reads as a machine's. And a question asked of the archive is answered in the church's terms rather than the dictionary's.

What to ask

Does the system learn between services at all, or does every Sunday start from zero?
What does it learn from — the transcript of our own services, or something else?
Who decides what reaches the live voice: a person, or the system?
Can we see what it has learned, correct it, and switch a thing off for good?
Does what it learned also reach the transcript, the summary and the drafts, or only the voice?
What happens to what it learned when we delete a service?

Tick what you have asked. It stays on this device.

How Sanctavo does it

Sanctavo keeps both. The glossary is a versioned list per language pair; a new version is staged and a live service can only read one the church's team has approved. Church Memory sits beside it: after each service the system notes names, places, ministries and phrases, shows them in the console with how many services they were seen in, and uses a learned item only once it has recurred in separate services. What the church typed always wins, learning can never overwrite it, an item can be pinned or switched off, and editing a learned item makes it the church's own. Both reach the live translation and every draft written afterwards. Erasing a service recounts the evidence. Because the system uses AI, errors remain possible; the memory exists so that the same one is not made twice.