Recaps

Before every reply, your writer gets sent the story so far. There's a hard ceiling on how much that can be, set by the model you picked and the context size you gave it.

A long evening blows past it. When that happens, mi·do·na takes the oldest stretch of turns, has the model write it up as a short summary, and sends that in place of the turns themselves. Those are recaps. They stack up into a running account of the story that rides along with every reply from then on.

You decide when it happens. mi·do·na tells you when there's enough piled up to be worth compressing, and recap now does it. Nothing gets summarised behind your back.

Recent turns stay exactly as written. A tenth of the context is held back for them and never touched, so the last stretch of story is always verbatim. Anything you've pinned is kept whole too.

Recap cadenceLess often, slower / Recommended / More often, faster

How much unsummarised story is allowed to pile up before mi·do·na asks you to recap. More often keeps the pile small and steady, at the cost of recapping more. Less often lets the story run untouched for longer, then compresses a bigger stretch in one go. This doesn't change how recapping works, it changes how long until we bug you about it, and how long you need to wait for it to finish. Frequent pings, faster runs. Rare pings, long runs.

Recap resolutionLess detail, less space / Recommended / More detail, more space

How much story goes into a single recap. Imagine you are told to recap a scene from a movie as one paragraph. You could fit more detail writing about that scene than you could if you needed to recap the first act of a movie, or the entire movie. The more plot you need to fit into a preallocated space, the more detail you need to cut. Recaps work the same way: High resolution = more detail, more paragraphs. Low resolution = less detail, fewer paragraphs.

Here's the part worth sitting with. Recaps aren't free storage. They're text, they take up room, and that room comes out of the same budget as your live story. Every recap you've ever written gets sent on every single turn. So turning resolution up remembers more and leaves you less space to play in, which brings the next recap around sooner.

No instinct either way? Leave both on Recommended.

Under the dropdowns is a token count with a bar beneath it. The count is how much story has piled up since your last recap. The bar is how close that pile is to being worth compressing.

That point moves with your setup, not just your cadence. A bigger context size in the Model panel fills the bar slower. Long personas and a thick stack of recaps fill it faster, because those get sent every turn too.

chat more first

There isn't enough piled up yet to make even one recap worth writing. Nothing to do.

recap now

There's enough to compress. Best hit at a natural break in the story rather than mid scene.

Recap due

The pile has hit the point your cadence sets. Good moment to hit recap now.

Context overflowing

The live story doesn't fit anymore, and turns are getting left out of what gets sent. Recap now.