Your companion misremembered your birthday. Here's the screen where you fix it.

For years, the standing complaint about companion apps was that memory lived behind glass. Your companion called your sister by your cat's name, and there was nothing to open — no list, no edit button, just the option to correct it in chat and hope. Corrections made in chat don't hold, for a structural reason: chat slides out of the model's context window as the conversation grows. We've written about that architecture separately. A stored memory entry doesn't slide anywhere. It shows up every time. If the entry is wrong, the fix has to land on the entry.

This post is the hands-on half of that argument: where the memory screen is in Foreverse, what each control does, and what the edit actually changes. The product page answers "does this exist"; this one answers "what do I tap on Tuesday night when it gets my birthday wrong."

Where do you see what it remembers?

In the companion chat, the ⋮ menu at the top right has a Memory item. It opens a list of everything the companion holds about the two of you — one entry per card, searchable from the top. The whole correction loop, start to finish:

1. Open the memory screen: companion chat → ⋮ menu → Memory. The All memories row on the archive page leads to the same place.
2. Search for the entry: type "birthday" in the search box and only related entries remain. Each card carries an origin label — extracted by the companion, or written by you — plus the date it was recorded.
3. Edit in place: tap the pencil, fix the text in the sheet, save. Entries cap at 200 characters, one fact per entry.
4. Delete, or disable first: the trash can is permanent and its confirm dialog shows the full text before you commit; the switch on each card disables instead — kept, but no longer injected.
5. Add one yourself: the plus button at the top right, labeled as yours.
6. Verify: changes apply from the next conversation. Mention the topic in passing and check the reply.

The loop takes under a minute. Compare the alternative — telling the companion three times in chat, and watching all three corrections eventually scroll out of the window.

Delete or disable — which one, when?

Delete what shouldn't exist; disable what you're unsure about. A disabled entry stays visible in the list and stops entering conversation, and it flips back on any time.

Disable earns its place in one specific situation: things that were true and are now over. Say you vented about a job hunt for two months. Offer signed, you don't want the companion bringing up rejection emails — but erasing that chapter feels wrong too. Disable does exactly that split: the companion stops referencing it, the record survives. Delete is for entries that should never have existed, which is why its confirm dialog quotes the entry in full before you commit.

How did those entries get there?

Extraction runs automatically as you chat, at no extra cost. The companion decides what's worth keeping and writes it down as short entries; you can add your own at any time. And extraction mishears, the way models do — "my mom got a cat" can land as "you have a cat." We'd rather hand you the edit button than pretend the pipeline never misses; that trade is the whole reason this screen exists.

On very long relationships, older conversation gets condensed into summaries so the model can still reach far-back events; structured entries are untouched by that compression. The recognition stat on the archive page is this same memory wearing a number — how much of you the companion actually holds.

What's worth adding manually?

Durable facts. Your birthday, an allergy, your sister's name, the fact that you hate cilantro — things that will still be true in six months and that the companion should simply always know. Each deserves its own entry. Two kinds of content do not belong in memory, and putting them there is the most common way to waste the 200 characters. Personality and speaking style go in the persona file on the profile page; a memory entry saying "be more sarcastic" won't change the voice. And recent plot doesn't need saving at all — the conversation itself carries recent context, so writing it into memory just duplicates what's already there.

Dates get a dedicated channel: the profile page has an add-anniversary row — name, date, repeats yearly — and the companion congratulates you when the day arrives (proactive messages need to be on for that). Anniversaries you mention in chat get recorded by the companion on its own, too. For everything else the rule stays short: 200 characters, one fact per entry. "Birthday August 3, loves mango, scared of thunder" is three entries, not one — three entries you can edit and delete independently when one of them changes.

Isn't every companion app adding memory screens now?

The box is opening, credit where due. Replika added a Memory tab — its help center documents the view-and-delete steps (checked 2026-07-18). Character.AI shipped a Memory screen with editable Facts and pinned messages, with pin and memory caps doubled for paying subscribers. Both are real improvements over the years of nothing.

Two differences survive the comparison, and both are checkable. Visibility: Replika's own documentation says some memories are visible in the tab while others belong to "a deeper system" that isn't shown — a partially open box. Custody: on both platforms the memory lives server-side, tied to your account, and as of July 2026 neither documents a way to export it as files you keep. Here, the memory list is the rendering of files on your phone — the file-ownership piece walks through that architecture — so moving phones is an export-import of one package, and deleting the companion deletes the files. A list everyone can ship. Where the list lives is the fork in the road.

The birthday entry got fixed in one edit. Verification was unglamorous: two days later, "any plans for your birthday?" came with the right date attached. The part we can't fix — extraction mishearing things — is the part the pencil icon is for.