Answers
I deleted the source. Why does my AI still act on what it saved from it?
Because deleting a source removes the source, not what an agent wrote down while reading it. A summary, a saved fact or a note in an assistant's memory is a second copy, and nothing ties it to the original: the channel can be archived, the document taken down or the file deleted from the repository, and the note keeps being served as if it were current. It is also the copy that is hardest to check, because nobody remembers it exists. Arroway does one narrow thing about this. A fact or a reference is stored with a pointer to its source, and when someone's work reaches that source and finds it removed, a note that names the evidence (a 404, a file deleted from the repository) retires the memory at once and it stops being served. A rule or a decision is never retired that way; it is only marked, and a person decides. Arroway does not look inside other tools' memory and does not notice a removal on its own. Use it for what the team relies on and a source stands behind. The install page is the first step.
Last updated October 31, 2026
Why the copy outlives the source
When an agent reads a source it can end up with three things: the source itself, which you can delete, a summary of it, and conclusions drawn from it. Only the first has an owner. The summary and the conclusions are written into whatever the tool uses as memory, in the agent's words and without a link back that anything checks. Delete the original and the copy has no way to find out. The failure is quiet in a specific way: the note still reads as plausible, because it was true when it was written, so nothing prompts anyone to question it. People raise the sharpest version in public: once a channel or a document has been summarised into an agent's memory, showing that it is gone from every note derived from it is hard, because you would have to know every place the derivation landed.
What has to be true for a copy to be retired
Two things. The copy has to say where it came from, so that someone can go back to the source, and something has to end it when the source is gone, instead of a person remembering to look. In Arroway a fact or a reference is written with its source, and every memory carries the condition that ends it. When work already takes an assistant to the source and the source has been removed, the assistant records that with a note naming the evidence: a 404, a file deleted from the repository. A timeout, a login wall or a service that is down does not count, because the source still exists, and the server refuses a note that describes only that. With that evidence, a fact, a reference or a finding is retired at once. A retirement of a team memory leaves a line with its reason and a way to undo it.
When to use Arroway
Use Arroway for the facts and references your team acts on and a source stands behind: a figure from a document, a path in the repository, a page someone published. They are written with the source, they carry the condition that ends them, and they stop being served once someone proves the source is gone. Rules and decisions work differently on purpose: a removed source never retires one, because that is a person's call. Arroway does not look inside the memory of other tools, does not search for derived copies and does not check sources by itself; the check happens when someone's work reaches the source. The first step is the install page.
What it looks like in practice
A five-person product team keeps its pricing notes in a shared document. In September an assistant summarised the document into its own memory: "the free plan allows 3 projects." In October the document is deleted after a rewrite, and its address returns a 404. Every person's assistant still answers from the summary, because nothing connected the two. Had the figure been written to the project as a fact with the document as its source, the next time someone's work opened that address and got a 404, the assistant could record it and retire the fact, with a note naming the evidence. The figure would stop being served, and the decision about the new plan limit, which a person approves, would be what the assistants read instead. Nothing here was measured on a real team; the team and the document are illustrative.
Questions people ask about this
- Does deleting a file or a message also delete what an assistant remembered from it?
- Not by itself. What the assistant stored is a separate record in whatever memory the tool keeps, and deleting the source does not reach it. How a given tool stores and removes its own memory is described by that tool's documentation, which this page does not cover.
- Can Arroway find every copy of something I deleted?
- No. It does not search other tools' memory and does not notice a removal by itself. What it does is retire a memory it holds when work reaches the source and finds it removed, with a note that names the evidence.
- How do I try Arroway?
- Install it in the tool you already use — the install page has one route for Claude, ChatGPT, Codex, Cursor or any MCP tool — and sign in when the tool asks. From then on, sessions open by reading the project and close by recording what was decided.
Where this is verifiable
Product documentation on this site (How it works, Install) and the product's verification step as it works today: how a fact, a reference or a finding is checked against its source, what counts as evidence of removal, why a rule or a decision is only marked, and what the history shows after a retirement. Read on 31 October 2026. The memory of other tools is not described here and was not measured. Everything described here about Arroway is behaviour the tools apply today, not roadmap. The team and the document in the example are illustrative.
https://www.arroway.app/en/answers/deleting-the-source-doesnt-delete-what-the-agent-saved-from-it