Guide

Finding the email you cannot remember the words for.

Your mailbox search matches characters. Your memory stores meaning. That gap is why you can picture a message clearly and still fail to find it, and closing it is the single most useful thing an AI does with a mailbox.

  • Reading About 5 minutes
  • Tested on 26,930 messages, 6.63 GB
  • Read flag Unchanged by searching

The short version

Do not guess the words. Describe what you remember, roughly when and roughly who, and let it run several literal searches for you. Name the likely folder first, because each folder searched is another call.

The problem

Search matches strings, you remember meaning

You remember a message the way you remember a conversation: it was the builder, some time in the spring, about the bit at the back, and there was a photo. Every one of those is meaning. None of them is a string that appears in the message.

So you type "builder invoice roof" and get nothing, because the message came from a company name you have forgotten, called the work a "quotation", and referred to guttering. Three misses out of three, and the message is right there.

What an AI adds, precisely

Not a better search. The same search, run many more times, with the guessing done by something that is good at guessing. It expands your one vague description into a dozen literal attempts, tries them, reads the summaries, and tells you which look right. That is a mechanical advantage rather than an intelligent one, and it is why it works.

Searching a large mailbox with an AI: the robot holding a lit lens over a deep drawer of filed letters, picking one out.
Many attempts, made quickly, instead of one good guess

Mechanics

What mailbox search actually does

Worth knowing because it explains every frustrating near-miss you have ever had. IMAP search, which is what runs behind the box in most mail clients, is a substring match over whole fields. Ask it for SUBJECT "roof" and it returns every message whose subject contains those four characters in that order, anywhere.

What you expect What actually happens
"invoice" finds "invoices" and "invoicing" It does, because they contain the substring. But "bill" and "quotation" are not found at all.
Results come back best first There is no ranking. Results come back in mailbox order and every match is equal.
A typo still finds it It does not. There is no fuzzy matching, so one wrong character returns nothing.
It searches everywhere at once Search is scoped to one folder. Your mail client may hide that by searching several in turn.
It looks inside the PDF It does not. Headers and body text only.

Gmail's box is not IMAP search

Google's own web search does have stemming and ranking, which is why searching in Gmail sometimes finds things that a mail client cannot. If you have been comparing the two and concluding your desktop client is broken, it is not: the two are doing genuinely different operations.

Walkthrough

The method: widen, then narrow

Backwards from how people usually go about it. The instinct is to start with the most specific query you can manage and loosen it when it fails, which means every attempt returns nothing and you learn nothing from any of them.

  1. Start with the axis you are most certain about. Usually the person, sometimes the month. Not the subject, because subjects are what people misremember most.
  2. Get a list, not an answer. Forty summaries you can scan is better than one confident wrong result. You will recognise the right one instantly.
  3. Narrow on what you see. The list will remind you of the thing you had forgotten, which is nearly always the company name.
  4. Open exactly one. Reading it does not mark it read, so opening the wrong one costs nothing at all.

The three axes worth naming

Who is the strongest, because senders are stable and you rarely misremember a person. When is next, and "spring last year" is a usable constraint even though it is vague. What is the weakest and the one to give last, because it is the one your memory has paraphrased.

Worked example

What to actually type

Describe the memory. Do not attempt the query.

The vague one, which is the point

I am looking for an email from some time in spring last year. It was from a building firm, I cannot remember the name. It was about work at the back of the house and I think there was a photo or a PDF attached. Search my Archive folder, try several different wordings, and show me a list of anything that might be it. Do not open any of them.

That gives it three constraints and permission to guess the fourth, which is exactly the division of labour that works. Then narrow:

The narrowing

Number four looks right. Open just that one and show me the body and what was attached.

Reviewing email search results with an AI: the robot laying out a short row of letters on a tray and pointing to one of them.
A list to recognise from, not one confident answer

Everybody has one message they gave up looking for. The free tier is enough for one honest attempt at yours.

Scope

Searching across folders

One search covers up to twenty-five folders in a single call. That changed on 27 August 2026: it used to be one folder per call, so asking about Inbox, Archive and three project folders was five calls and is now one. On the free tier of 5 a day, that is the difference between one question and most of your allowance.

Naming the likely folder first is still the better habit, and not for the old reason. Results are capped, so a search spread across everything can come back full and still be full of the wrong folder. In order of how often the answer is there:

  • Archive, if you archive rather than delete. Most old mail is here and most people forget it.
  • Inbox, if you do not.
  • Sent, which is the one people never think of and which very often has it, because you replied.
  • Trash, worth including before concluding it is gone.

Folder names differ by host

Some IMAP hosts prefix everything, so your sent mail is INBOX.Sent rather than Sent. Nothing is wrong if you see that. Ask for the folder by the name your own mail client shows you, and if you are unsure, ask for the folder list first.

Naming the likely mail folder first: the robot pointing into one open drawer of a four-drawer filing cabinet.
Name the likely folder before widening the search

Connect a mailbox and describe the one you cannot find. It is the fastest way to see whether this works on your memory.

Limits

What it will not find

  • Inside an attachment

    Anything inside an attachment. A PDF, a spreadsheet or an image is not searched, only the message around it.

  • Emptied from Trash

    Anything already deleted and emptied from Trash. Nothing here recovers mail.

  • Another mailbox

    Mail in a different mailbox. One connection is one mailbox, and shared or delegated mailboxes are not supported.

  • Folders not searched

    Anything in a folder outside the ones a search covers. One call reaches up to twenty-five of them, not the whole account by default.

  • Misremembered entirely

    A message you are misremembering entirely. It is good at finding things you half remember and no use at all on things that were never there.

How do I find an old email I can barely remember?

By describing what you remember instead of guessing the words. Tell it the shape of the thing: roughly when, roughly who, what it was about, what was attached. It turns that into several different literal searches, runs them, and reads the summaries back. You are outsourcing the guessing, which is the part you are bad at and it is not.

Why does searching my mailbox miss things?

Because mailbox search matches strings, not meaning. There is no stemming, no synonyms and no ranking: a search for "invoice" does not find a message that says "quotation", and a search for "roof" does not find one about guttering. Your memory stores the meaning and the search box wants the exact characters, and nothing bridges that gap on its own.

Can an AI search across folders?

Yes, and up to twenty-five folders in a single call. That changed on 27 August 2026; before it, four folders meant four calls. Naming the likely folder first is still the better habit, because results are capped, so a search spread across everything can come back full of the wrong folder. Ask for Inbox, then Archive, then Sent, in that order.

How big a mailbox can it work through?

The test mailbox holds 26,930 messages across 6.63 GB and was read in a single session over IMAP. Size is not the limiting factor; the daily call limit is, because each search is one call however many folders it covers.

Can it search inside attachments?

No. Search covers headers and message bodies. A PDF attached to a message is not indexed, so a message is findable by its covering text and not by what the document says. If you need the contents of attachments searchable, you need a document system rather than a mailbox.

Does searching mark anything as read?

No. Search returns summaries in the same shape a listing does, so nothing is opened. Even reading a result in full does not mark it read, because that is a separate action.

Attribution

Sources

Keep going

Read next

Triaging a full inbox with an AI, without losing anything

Read-only first, decide in batches, act last. The method leans on one property of the engine: reading a message does not mark it read, so a triage pass you abandon halfway leaves no trace.

Folders, filing, and the mail an AI should not touch

What a move actually does to a message, why delete means Trash and not gone, how folder naming differs by host, and the categories worth keeping an AI away from entirely.

What an email MCP server is, and what one can actually do

The protocol in plain terms, the twenty-eight tools a mail server exposes, and the difference between an AI that can describe your inbox and one that can work it.

Try it on the one you gave up on.

Everybody has a message they stopped looking for. The free tier is 5 calls a day, which is enough for one honest attempt at it.