How memory works
A conversation and a memory are not the same thing — and that difference is the one thing people ask about most. Here is exactly how memory is built, used, and forgotten.
A conversation is not a memory
A conversation is the literal thread of messages you're reading right now — kept until you delete it, or your retention setting expires it automatically. A memory is different: a short, standalone fact your companion wrote down, like “Mia started watercolour classes in March,” so it still knows that once the conversation has scrolled out of reach. Conversations are raw and temporary; memories are the durable layer underneath them.
How a memory gets created
After a conversation, a second model call reads back through it and proposes short structured memories worth keeping — the same one that noticed Mia's watercolour classes might also note that you mentioned feeling nervous about the first one. Every memory records which message it came from, so the Memory browser can jump straight back to the moment it was learned.
The categories
Every memory is filed under one of seven categories: life, people, interests, patterns, history, goals, or care. Watercolour classes would land under interests; a note about your sister's new job would land under people.
How your companion picks which memories to use
Your companion is only ever handed around 40 memories per turn — the ones most likely to matter right now. Each memory is scored as 0.15 × importance + 0.60 × recency + 1.20 × relevance, so a fresh, closely relevant memory outweighs an old, only loosely related one. Relevance is meaning-based, not keyword matching — it recognises that “the watercolour class” and “painting lessons” are the same thing. Anything you've pinned is always included first, no scoring needed.
Deduplication
If you mention the watercolour classes again a week later, the companion doesn't file a second, near-identical memory — a new memory that closely paraphrases an existing one in the same category updates that one instead. A nightly job also sweeps for near-duplicates the write-time check missed, using a stricter threshold, and archives the loser rather than deleting it outright, so it's still recoverable.
Decay, the 400-memory cap, and what “archived” means
Memories that sit unpinned and untouched for more than 90 days lose a single point of importance — at most once per 90-day window, and never below 1 — so old, unremarkable details naturally fade in priority rather than staying loud forever. Each account holds at most 400 active memories; past that, the lowest-ranked ones are archived to make room, and pinned memories are effectively exempt from all of this. Archived isn't deleted — it just stops being surfaced.
Your conversation retention setting
Separately from memory, you can set how long conversations themselves are kept. If you set a limit, a nightly job deletes conversations older than that automatically. This setting only ever touches conversations — memories already extracted from them are untouched, and rolling summaries of long conversations (made every 20 messages) mean your companion keeps useful earlier context even after the raw thread is gone.
Your controls
Every memory your companion holds is listed in your Memory browser. Edit it if it got something wrong, pin it so it's never dropped, archive it, or delete it permanently. Export everything as a single JSON file from Settings whenever you like. Delete all your memories, or your entire account, at any time — both are immediate and permanent.
When memory is wrong
A memory can be misheard, out of date, or just misremembered — the same limits that apply to the model writing your replies apply here. If your companion says something about you that isn't right, open the Memory browser, find the memory, and correct or delete it directly; there's no need to argue with it in conversation.
Privacy, and how the model itself works
For where memories are stored and who can see them, see Privacy, plainly. For what happens to a message before it ever becomes a memory, see How the AI works.
Questions? Email hello@myaifamilyfriend.com.