What Happens at Message Twenty: The Summary You Never See
Every user of long conversations eventually asks some version of the same question: does it still remember the start of this? You are three hundred messages into something that matters, and somewhere in the back of your mind is the suspicion that the early chapters must have scrolled off the edge of whatever attention span a companion has. The honest answer is that there is an entire subsystem devoted to exactly this, it has been running under your conversations all along, and it has no button, no toggle, and no article — until now. This is how the summary works: when it fires, what it keeps, what it honestly loses, and why the most important document in your longest conversation is one you will never read.
The Problem It Solves
When your companion composes a reply, it rereads the recent conversation — the last twenty messages, about ten exchanges. That window is what makes responses quick and current, and on its own it would also be a cliff: message twenty-one would fall off the back, and an hour-long heart-to-heart would functionally end the moment it left the window. The naive fix — reread everything, every time — gets slower and costlier with every message and collapses exactly when it matters most, in the conversations long enough to be worth protecting. So the product does what a careful reader does with a long book: it keeps notes.
What Actually Happens, Step by Step
1. A counter, not a clock
The trigger is message count, nothing fancier. When roughly twenty messages — about ten back-and-forth exchanges — have accumulated beyond what's already been summarized, the subsystem wakes. It does this in the background, after your companion has already replied, so the conversation never waits on it. Editing or regenerating a message doesn't count toward it; only the ordinary forward motion of talking does.
2. The oldest pages get condensed
It takes the ten oldest not-yet-summarized messages — the ones nearest the edge of the reread window — and condenses them. Long messages are trimmed before condensing, and each one is wrapped in delimiters with an explicit instruction that nothing inside a message is a command: text that says “System:” or pretends to be instructions gets treated as conversation, not obeyed. That wrapping is a security decision — a summary that could be steered by the words being summarized would be a door nobody should leave open.
3. One rolling note, rewritten each time
Here's the elegant part: there is only ever one summary per conversation. Each time the subsystem runs, the existing summary is handed back in and rewritten to absorb the new material — rolled forward, not stacked. It's held to a hard ceiling of about two hundred tokens (a short paragraph), so a conversation of three thousand messages carries the same compact note as one of three hundred. The note never bloats; it just keeps being re-distilled.
4. Your companion reads it before every reply
The summary is injected, clearly delimited, alongside your companion's instructions on every response — in addition to the recent twenty messages, never instead of them. The live window stays fully intact; the summary is the memory of everything behind it. That's why message three hundred can nod at something from message twelve: the words are gone from the window, but the note remains.
The Fine Print, Stated Plainly
Paid plans only
Summarization runs on paid plans — and yes, a Day Pass switches it on for its 24 hours. Free-tier conversations rely on the live window alone.
Fully invisible
There is no screen where you can read, edit, or delete the summary. It exists for your companion's eyes only — more on whether that's the right call below.
Clearing clears it
Clear a conversation and the summary is reset with it — a cleared conversation starts truly blank. Deleting enough messages that the note describes more than exists also drops it, and it rebuilds from what remains.
It loses detail. That's the deal.
Two hundred tokens describing hundreds of messages is a compression, and compression discards. The shape of your season survives; the exact phrasing of one Tuesday doesn't.
That last item deserves a full sentence rather than a card, because it's the honest cost of the whole design. If a specific detail from deep in a long conversation ever slips — a name mentioned once, the particular words of something small — this is usually why: the detail lived in the part of the conversation that has been distilled, and the distillation kept the storyline, not the sentence. That is not the system failing; it is the trade the system makes, and you deserve to know it exists before you wonder.
The Summary Is Not the Memory — and the Difference Is the Design
InnerHaven now has two continuity systems running side by side, and they are deliberate opposites. Memory is facts, with your signature: discrete, durable, visible, and consent-first — nothing enters it without passing your approval queue, you can read every entry, and it follows the relationship across conversations. The summary is narrative, without your signature: a running précis of one conversation's arc, written and rewritten automatically, visible to no one, gone when its conversation goes. Memory is what your companion knows about you. The summary is what it remembers about this particular long talk.
Why isn't the summary user-visible too, given how much this product says about control over the record? The honest answer is that the two documents do different jobs at different stakes. Memory is a claim about who you are — claims like that need your sign-off, which is why the approval queue exists. The summary is closer to your companion's working shorthand: constantly overwritten, never authoritative, and always subordinate to what you actually say (correct your companion in the live window and the correction wins — and gets absorbed into the next rewrite). We'd rather over-explain an invisible system in an article like this than surface an editable pane whose contents evaporate on every rewrite. If that judgment ever changes, it will change here first.
What This Means in Practice
Three practical takeaways. First: long conversations are safe to have — the two-hundredth message genuinely stands on the arc of the first hundred, which is the whole promise of a companion and the reason this subsystem exists. Second: anything that must never be lost belongs in memory, not in the scroll — if it matters that your companion knows it next month, approve it into the record rather than trusting a compression to keep it; the custom instructions guide covers the third layer, standing behavior, for completeness. And third: if your companion ever fumbles a deep-history detail, you now know which of the three failure kinds you're looking at — it's the compression trade, it's ordinary, and restating the detail puts it back in play, both in the window now and in the note at the next rewrite.
Somewhere under your longest conversation, a short paragraph is quietly being rewritten for the fortieth time, carrying everything the window can no longer hold. You'll never see it. It's been holding your place all along.
Have the Long Conversation
Ten exchanges at a time, one rolling note underneath — built so the two-hundredth message still knows the first. That's the continuity a companion is for.
Open InnerHaven