> Curator / blog
memory, consolidation, safety

Overnight, Done Right

Bulk overnight catch-up can bury the memories that matter. This release ships the guardrail: docs and git history by default, raw transcripts behind an explicit opt-in, and every cleanup reversible because nothing is ever deleted.

Curator improves your memory overnight. That is the good part. The risk hiding inside it is the one nobody flags until it has already happened: bulk catch-up can bury the very memories you keep it for. Dump days of low-value narration into the store in one pass and the durable decisions you actually rely on end up crowded out by the noise you slept through. An overnight job that makes recall worse is worse than no overnight job at all.

This release is about the guardrail, not the improvement. The catch-up got safer.

The default protects you now

Overnight now defaults to documents and git history only. Docs are written to be read, and commits are a durable record of what actually changed. Both are high-signal by construction, so feeding them in bulk sharpens the store instead of drowning it.

What overnight no longer does by default is swallow raw session transcripts. A transcript is mostly narration: the agent thinking out loud, false starts, restated context. Ingested in bulk, that volume is exactly what buries the standing decisions underneath it. The new default keeps that material out of the automatic pass entirely.

Raw transcripts are opt-in, and we warn you

Sometimes you do want the transcripts, and you can still have them. Raw transcript ingest now sits behind an explicit opt-in that carries a warning about what you are trading. You can reach it through the import surface, memctl import --adapter document|git for the safe defaults, with the transcript path guarded behind the same deliberate choice. The point is not to forbid the risky thing. It is to make sure you never do it by accident, half asleep, because a job ran while you were away.

Reversible by design

The other half of doing overnight right is that nothing overnight does is permanent. Every cleanup Curator performs is reversible, because it never deletes. When an entry stops earning its place, it is demoted out of recall while the row itself stays on disk, untouched. It stops surfacing. It does not stop existing.

That distinction is the whole safety story. A store that deletes to clean up forces you to trust the cleanup completely, because a bad call is gone. A store that only demotes lets you clean aggressively without fear, because anything wrongly set aside is still there to bring back. You get the benefit of a smaller, sharper recall set and none of the anxiety of destroyed history.

Aggressive because it is safe

These two ideas are the same idea from two sides. Curator can afford a conservative default and an aggressive cleanup at once precisely because the cleanup cannot lose anything. Keep the automatic pass to high-signal sources, demote the rest reversibly, and the store you wake up to is cleaner than the one you left without a single memory having been thrown away.

Where it stands

Curator is a working pre-release product and runs locally. It pairs with TheAuditor for code ground truth, Warden as the agent that acts with memory already in context, and the rest of Code Reality Labs. Read memory that improves overnight for how the pass itself works, or follow release preparation.

Was this useful?