Whizi memory: what it remembers across your chats

The short answer

Yes, Whizi remembers durable facts you share, and reuses them in later conversations. Settings describes it in one line: "Whizi remembers durable facts you share in conversations, like your name, preferences, and goals. It uses them across your chats on web and mobile."

The facts belong to your account, not to a single conversation, so they follow you into a brand new chat and into whichever model you pick. They are stored as a plain list of short bullet lines, capped at 2,000 tokens.

Surface	What memory does there
Text chat	The saved list is attached to the request as its own system message before the model answers
Voice	The saved list is added to the voice session prompt when the session starts
Settings	A "What Whizi remembers" row lists the facts, and one action clears all of them
Data export	The export file carries the list under a memory field
Account deletion	Saved memory is deleted along with the rest of the account

What earns a remembered fact

The test a fact has to pass is strict: it qualifies only if it would change how Whizi answers weeks from now, in an unrelated conversation. Mentioning something is not asking to be remembered, and returning the list unchanged is the normal outcome of a pass.

Recorded	Not recorded
Durable facts you share about yourself: name, preferences, goals, relationships, ongoing situations	Things you merely asked about or discussed, small talk, one-off tasks, transient state, passing opinions
Standing preferences you explicitly ask to be remembered	A joke, an experiment, or a one-off request treated as a standing rule
The newest version when two facts conflict	Health or financial details, unless you explicitly asked for them to be remembered
Names, numbers and dates kept verbatim	Facts about Whizi itself: its name, interface, features or models, or what you believe about them
Facts folded into an existing line where they fit	Anything a line you already have implies

Whizi is instructed never to store observations about the app itself, and never to store instructions about how the assistant should speak or format replies unless you explicitly asked for that as a standing preference. A stored rule of that kind would be obeyed in every future conversation as if you had just asked for it, which is exactly the failure the rule exists to prevent.

Voice gets one extra guard. A voice transcript comes from speech recognition and can contain mishearings, so a line that would read as garbled or surprising is left out rather than guessed at.

Memory is not the same thing as per-project instructions. Instructions you want applied to one body of work belong in a project, covered in use projects.

When Whizi writes to memory

Not after every message. In text chat the extraction pass runs when the conversation crosses a multiple of six turns, which works out to roughly every third reply. The pass reads the last six turns plus the list you already have, and returns the rewritten list.

There is one deliberate bypass. If your message asks to remember something, using wording like "remember", "memorize", "do not forget", or asking Whizi to save or keep something in memory, a pass fires immediately rather than waiting for the next boundary. That exists because a first exchange of "my name is..." followed by "remember that" used to extract nothing at all, since a conversation's first scheduled pass only arrives at six turns.

Voice works on a different trigger: the pass runs once at the end of a session, over that session transcript.

Every pass runs in the background after your reply is already on screen, and it rewrites the whole list rather than appending to it. That is also how stored junk gets cleaned up: existing lines are held to the same bar on every pass, and a line that breaks the rules is deleted rather than kept out of caution.

The read rides alongside the rest of the request rather than in front of it, so a memory failure never blocks a reply. If the read fails, the model simply gets no memory block for that message.

The size ceiling

Memory is capped, which is why it stays a short list of dense lines rather than a growing transcript of everything you ever said.

Setting	Value
Maximum stored memory	2,000 tokens
Size a rewrite aims at	1,500 tokens
Shape it aims at	About 50 lines, about 55 characters each
Chat turns read per pass	The last 6
Transcript input cap, chat and voice	20,000 tokens
Input cap on either model pass	50,000 tokens

When a rewrite comes back over the ceiling, a second pass shortens it: overlapping lines about the same person, thing or situation are merged, and the least useful lines are dropped. Usefulness is judged on content, never on position in the list, and long standing facts about who you are (name, family, work, standing preferences) are treated as the most valuable lines rather than the most disposable.

A positional trim survives underneath that as a last resort floor, for the case where the shortening pass fails or comes back still over budget. Memory has to come back under the ceiling because the chat system prompt is budgeted around it.

Seeing and clearing what Whizi remembers

On the web, open Settings. The Memory section sits at the end of Preferences, with a row labelled "What Whizi remembers" carrying a count of the stored facts. Opening the row expands the full list, with a red "Clear memory" action underneath it.

Clearing asks first. The confirmation is titled "Clear Whizi's memory?" and reads: "This permanently deletes everything Whizi remembers about you, across web and mobile. This can't be undone."

With nothing stored yet, the row is replaced by "Nothing remembered yet. Facts you share in chats will show up here."

In the app, the same thing lives in Settings under the row "What Whizi knows about me". It opens an alert titled "What Whizi remembers" with the facts as bullets, a Close button, and a destructive "Forget everything" button. A successful clear confirms with "Memory cleared" and "Whizi no longer remembers anything about you."

There is no way to edit or delete one individual fact, and no switch that turns memory off.

When memory does not load

Both memory failure strings name the memory service, not your account.

What you see	What it means
"Memory is unavailable right now. Please try again later."	The Settings read failed. Your stored facts are not affected
"Failed to clear memory. Please try again."	The clear request failed, so the list is still there. Retry the clear

In the app the equivalents are alerts titled "Memory unavailable" and "Memory could not be cleared". Neither one deletes anything, so a failed clear leaves your facts exactly as they were.

Guests, privacy and getting the data out

Memory works without an account. A guest session gets its own fact list, and when that guest creates an account the facts are merged into the account line by line, without duplicating any line, and the guest copy is deleted.

Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models, and does not sell that content as training data. The full detail is in what Whizi does with your data.

Memory is the one exception to picking your own model. The background passes that write and condense your fact list do not run on the model you selected for the conversation. They run on a fixed model Whizi picks for that job, and what they read is the recent turns of the conversation plus the fact list you already have. Switching your chat to a different model does not change which model performs the memory pass.

Saved memory is kept until deleted, unlike uploads and generated media which are configured to expire after up to 30 days. Where each category physically lives is set out in where Whizi stores your data.

Your data export includes the list, under a memory field alongside your chats, projects, media, share links, subscription and referral profile. The account data export is covered on the Whizi API page. Account deletion removes saved memory along with the account itself.