How to find a reel you saved six months ago
A practical retrieval playbook: what you actually remember about an old save, which of those memories are searchable, and how to get from a vague recollection to the right video.
You almost never go looking for a saved reel with a clean description in hand. What you have is a fragment. A man in a workshop said something about letting the glue cure overnight. There was a lot of orange in it. It was around the time you were flat-hunting, so probably March. You would recognise it instantly if you saw it, and you cannot describe it well enough to find it.
That is the honest starting condition, and it is the reason "just search for it" so often fails. A search box can only match on text that exists somewhere in the system. Your fragment might be a memory of something that was never written down at all. So the first move in retrieval is not searching — it is working out which parts of what you remember are the kind of thing a machine could have recorded.
Sort your memory into searchable and unsearchable
Take the fragments you have and split them. Some of them are text, or were converted into text at some point in the pipeline. The rest are impressions, and impressions do not go into a query box.
- A phrase somebody said out loud — searchable, if the audio was transcribed. This is usually your strongest lead, because spoken sentences carry the actual substance of the video.
- Text burned into the frames — searchable, if the frames were read with OCR. Ingredient lists, sets and reps, tool names and prices tend to live here rather than in the audio.
- The caption under the post — searchable, and reliably so, because the creator typed it as text in the first place.
- The topic — searchable, if the reel was categorized. Not precise enough to find one video on its own, but excellent for cutting a library down to a scannable slice.
- The look of the thumbnail — not searchable. Colour, framing and the fact that it was shot in a kitchen are nowhere in the index.
- The vibe of the account — not searchable. "One of those deadpan Australian guys" is a real memory and a useless query.
Notice how much of a typical recollection lands in the bottom two lines. That is not a flaw in your memory; visual and tonal impressions are simply what video is best at leaving behind. The trick is to stop trying to spend them and to spend the top four instead.
Start with the rarest word you remember
Search rewards low-frequency terms, and it punishes obvious ones. This is the single highest-leverage thing to understand about finding an old save, and most people get it exactly backwards — they start with the most confident, most central word they can think of, which is almost always the word that appears in everything.
Say you are hunting a Korean fried chicken reel in a library with two hundred food saves. Searching "chicken" returns forty results and you are back to scanning thumbnails. Searching "recipe" returns most of the food library, which is worse than not searching. But searching "gochujang" returns three, maybe one. The word is rare, so it is discriminating. Same for "mandolin" over "slice", "flashing" over "roof", "keyset" over "database", "tenancy deposit" over "renting".
So the question to ask yourself is not "what was this about" but "what is the weirdest word that was probably said in it". Brand names, proper nouns, technical jargon, specific quantities, unusual place names. If you can produce one rare word, you are usually one query from done.
Then narrow by category, not by date
When no rare word comes to mind, the instinct is to reach for time — to scroll back to roughly when you think you saved it. That instinct is worth resisting. People recall dates terribly. Anchoring a save to "around March" usually means anywhere in a three-month window, and you will not know when you have scrolled past it, because a thumbnail you do not recognise looks exactly like a thumbnail that is not the one.
Topic is recalled far better. You may not know when you saved the chicken reel, but you are certain it was food, and you are certain it was not fitness or travel. Filtering to a category and scanning that slice is nearly always faster than paging chronologically through everything, because it cuts the candidate set by an order of magnitude on a memory you actually trust. Use the date as a tiebreaker inside a category, never as the first filter.
The four-step search
Put together, the whole procedure is short. Work down it in order and stop as soon as you have the video.
Pick the rarest word or phrase you can recall
Not the topic word — the odd one. A brand, an ingredient, a tool, a number, a name. If two candidates come to mind, start with the one you would be least likely to hear in any other video you have saved.
Search it across caption, transcript and on-screen text at once
You usually do not know which of the three signals carried the word, and you do not need to. A search that spans all of them means a word spoken but never written still finds the reel, and so does a word written on screen but never spoken.
If nothing matches, drop to the category and scan
Switch from precision to recall. Filter to the topic you are confident about, then read down the titles and captions rather than looking at pictures. Text scans fast; thumbnails do not.
If still nothing, widen to a stem of the word
Search "gochu" instead of "gochujang", "mandol" instead of "mandolin". Transcripts mangle proper nouns and loanwords constantly, and a partial match will catch a misspelling that an exact match throws away.
Why some searches come up empty anyway
Sometimes you do everything right and the reel does not surface. It is worth knowing the real reasons, because each one suggests a different next move.
Speech-to-text is confident and wrong in predictable places. Names of people, brands and dishes are exactly the rare words you want to search for, and they are exactly the words a transcription model has least training data for. A model that has never seen "gochujang" written down will produce something phonetically plausible instead, and your exact-match query sails straight past it. That is what step four is for.
Then there are reels with nothing to index. A silent clip with a music bed, no narration and no on-screen text has produced no transcript and no OCR output. All that exists is the caption, and if the caption was three emoji then the reel is effectively invisible to search. No query will rescue it; only browsing by category will.
And in Reel Mind specifically there is a mechanical limit worth naming. The dashboard filters the reels it has already paged into its local cache rather than querying the entire server-side library. On a small library that distinction never shows up. On a large one it does: if a match lives in reels that have not been loaded yet, you need to scroll far enough to pull them in before the search can see them. If you are confident a reel exists and search says otherwise, scroll first, then search again.
Retrieval is not really a memory problem. You remember plenty about that reel — you remember the orange, the workshop, the tone of voice. The gap is that almost none of it was ever written down, so almost none of it can be queried. Search the fragments that survived as text, use the rarest of them first, and fall back to browsing a category rather than the whole pile. That is most of the skill, and the rest is giving the next search something rare to grab onto.