Four tempos
Treating learning as one block and automating the whole block was the real cause of the over-design.
This is the real reason the original plan was over-designed. It treated “learning” as one thing and tried to automate the whole thing.
Split apart, there are four tempos, and how much automation each can take is different in every case.
| Tempo | When | Automation | Why |
|---|---|---|---|
| Capture | continuously, in-session | by hand | The only point where new information enters |
| Consolidate | weekly | human-approved | It involves deciding what to delete |
| Apply | every session | fully automatic | Safe because no judgement is involved |
| Watch | weekly | fully automatic | Changes nothing, only reports |
The human only has to hold the places where judgement lives. The other three can be automatic without dying quietly — because the fourth one is watching.
Why capture isn’t automated
It’s the place you most want to automate. Summarise at the end of every session and drop it into memory.
That isn’t what happens. When a session ends, candidate notes are appended unprocessed. No summarising, no judging. No model is called, so the cost is zero.
The reason is the lesson from the dead channel. Make capture an automatic summary and nobody ever looks at what the summary left out. And that only surfaces months later, when the thing you needed isn’t there.
In the current structure, summarising happens when I write the memory. It takes hands, so it stays in view.
Why only consolidation has a gate
Of the four, exactly one has an approval gate. The other three don’t.
The criterion is reversibility.
- Capture only appends, so deleting undoes it
- Apply is a read; it changes nothing
- Watch only sends a notification
- Consolidation deletes. And a deleted note has no way back
Deciding what to delete is a judgement, and it’s the kind of judgement where there is no way to learn it was wrong. So that’s the one a human looks at.
Why apply is fully automatic
Injecting memory into every session is automatic, because there’s no judgement in it — files get read and inserted, and what gets inserted was already decided.
There’s a lesson in that. “Can this be automated” is decided by whether judgement is involved, not by frequency. Apply runs dozens of times a day and is fully automatic; consolidation runs weekly and has a gate. Wanting to automate the frequent thing is natural, but frequency isn’t the criterion.
The watcher is what protects the other three
The watcher fixes nothing. It reads and reports.
But its existence is what lets the other three be automatic. The reason the previous machine could stay dead for months was that producing nothing had no consequence, and the watcher plugs exactly that hole.
Multi-machine handling lives here too. It’s registered on both machines but only runs on the designated one. Otherwise the same notification arrives twice, and a notification that arrives twice quickly becomes one nobody reads.