Writing with an agent

What 'manuscript memory' actually means

August 16, 2026 · The StoryProp team

'It remembers your whole story' is the claim every AI writing tool now makes, and it is worth being precise about, because the phrase covers four different capabilities — and a tool can have any one of them without the others. A writer weighing memory claims needs the anatomy, not the slogan. Here it is, layer by layer, as we build it in StoryProp — along with a worked example of memory actually doing something at drafting time, and the questions that separate the real thing from a search box with good manners.

1. The permanent conversation

The base layer is the project conversation itself: permanent and ordered, not a window that slides. It isn't summarised down as it grows, so a decision from March is still there in August, in the wording it was made in. Summaries and indexes can help find things, but they supplement the record rather than replacing it — the difference between an index and a paraphrase you're forced to trust.

This is where conversation-based tools — general chatbots such as ChatGPT, Claude, or Gemini — are structurally weakest for long work: the conversation is the memory, so the window is the ceiling. Anthropic publishes what happens at that ceiling. Its compaction feature, currently in beta, fires by default at 150,000 input tokens: the API writes a summary and drops every content block before it from subsequent requests. Its help centre says the consumer product summarises earlier messages to make room for new content. OpenAI documents only that prompt plus output must fit inside the maximum context length; neither it nor Google publishes what a chat does when it crosses that line, so this is Anthropic's documented behaviour rather than a law of the category. The mechanics are their own subject, and we've written about why chatbots lose the plot; the consequence for this taxonomy is simply that if the base layer becomes a summary, everything stacked on it inherits the summary.

2. Authoritative records

Raw history isn't enough, because a transcript treats your rejected ending and your accepted one identically. So the load-bearing facts get promoted into records: canon, characters, the story's shape, the style rules you confirmed. Promotion is deliberate — 'make this canon' is an instruction, not an inference — which is what keeps a brainstorm from calcifying into fact. What you agree to gets written down; what you don't stays a suggestion. And the agent reads these records before drafting, so being remembered actually changes what gets written.

It's worth watching that consultation happen, because it's the part marketing language skips. Take the demo manuscript we use around here, The Quiet House. Among its canon records is one about an object: the chipped cup on Mara's kitchen windowsill belonged to her mother; Daniel chipped it the night he left; and in nine years neither of them has said so aloud. Now the writer directs, in plain language: draft the moment Mara starts packing up the kitchen. Nothing in that instruction mentions the cup.

An agent that reads records before writing doesn't need it mentioned. The draft comes back with Mara wrapping the ordinary dishes in newspaper, quick and businesslike, and stalling at the windowsill — the cup last, held a moment too long, carried out in her coat pocket instead of the box. The writer never asked for that beat; the record made it available. And if a later scene has Daniel joking openly about the night he chipped it, the agent flags the collision — that silence was canon — rather than smoothing it over, because sometimes the contradiction is the reveal you were building toward, and sometimes it's a slip. Surfacing instead of silently correcting keeps that call where it belongs: with you.

3. Every draft, kept

Manuscript memory includes the versions. Each revision is retained, comparable side by side, restorable without destroying what came after; a whole agent pass can be undone as one action. The practical effect is courage: you can accept a bold revision knowing the cautious one still exists somewhere real, and you can answer 'which version of this scene was better' by looking instead of mourning.

4. Research with provenance

The subtlest layer. When the agent researches — a named book, a historical detail, a current fact — what it learned is saved with its sources and retrieval dates, and kept distinct from your story's truth. Claims stay attached to where they came from; a source's opinion never flattens into fact; and none of it contaminates canon unless you decide it should. Evidence and decision are different kinds of memory, and a tool that can't tell them apart will eventually cite your own invention back to you as research.

Memory is not search

It's tempting to collapse all of the above into 'good search over your project,' and the distinction deserves to be made plainly. Search answers questions you ask. Memory shapes work you didn't know to ask about. A tool can have flawless search and no memory at all: every fact in the project retrievable, and none of it consulted when the drafting happens. The burden of remembering-to-look stays with the writer — which is precisely the burden the tool was supposed to carry.

The packing scene shows the gap. While directing that scene, the writer wasn't thinking about the cup, so no query would ever have been issued. Search only helps once you know something is findable and what to call it, and at drafting time you mostly don't, because a novel accumulates more decisions than any writer holds in mind at once. There's a second difference, quieter but as important: search returns whatever matches, and your rejected ending matches the query too. Memory has to weigh authority — which is the whole point of promoting decisions into records instead of trusting the transcript to sort itself.

Four questions to ask about manuscript memory

The anatomy turns vague marketing into things you can check from the buyer's chair. First: can it retrieve the exact decision, or only the gist? Pick something you settled early — a name, a rule your ending must keep — and ask for the decision as it was made. A paraphrase that sounds right is the tell; gist is what compression leaves behind, and fiction runs on specifics.

Second: does remembering change what it drafts? Settle a fact, then sessions later ask for a scene that touches it without mentioning it — the cup test. Honored unprompted is memory; honored only when reminded is search wearing memory's clothes. Third: are the old drafts really there? Ask for the earlier version of a paragraph you revised and expect the actual text, viewable beside the current one, restorable without wrecking what came after — not a summary of what it used to say.

Fourth: does research come with receipts? Ask where a factual claim in your notes came from and expect a source and a date, not confidence with nothing under it. Choosing a tool for a long project involves more than memory, but memory is the part you can't retrofit later — and a tool ought to be willing to answer all four questions in public, which is what this post is. You can see how the layers fit together across the full feature set.

Memory sounds like a convenience feature. For long-form writing it's the difference between a tool that helps with sentences and a partner that can help with a book — because a book, structurally, is memory: promises kept across four hundred pages, from one chapter to another and from writer to reader. The tool that holds them with you is the one that gets you to the last page.

These four layers are the architecture StoryProp is built on — the permanent conversation, deliberate records, every draft kept, research with its receipts. It's in paid early access; bring a story with some history and ask it the four questions.

← All posts