Documentation
Memory and personalisation
What is stored, what triggers a save, what it is told to avoid, and how to delete it.
Memory carries durable facts about you between conversations: the courses you are taking, deadlines, work in progress, how you like answers written, and mistakes that keep recurring. It exists so you are not re-explaining your situation every session.
What triggers a save#
Capture is deliberately hard to trigger, and it always runs after the reply is delivered — it never sits between you and an answer.
- Message length
- Between 25 and 4,000 characters. Shorter than that is not a durable statement.
- Shape
- The message has to plausibly state something about you — first-person phrasing, a course, a goal, a habit, a preference. A one-off factual question never triggers it.
- Throttle
- At most one capture attempt every 3 minutes.
- Yield
- 0 to 3 facts per attempt. Near-duplicates of something already stored are dropped rather than stacked.
- Ceiling
- 40 memories. Over that, the least important and oldest are evicted first.
- Recall
- The 12 highest-relevance memories are placed in the system prompt of ordinary chats.
Answers to questions chud asked you#
When chud asks a clarifying question and you tap an answer, that pair is stored directly — no extraction call, no length gate. Up to three per turn, each 60 characters or shorter. A question and its answer are durable by construction: the model only asks when the answer changes what it produces. Without this, the single most reusable thing you ever tell it was the one thing memory could not keep, and you would be answering the same question every session.
This path currently ignores the memory switch
Turning off Remember things about me stops the extraction described above. It does not stop this one. Tapped clarification answers are still written to memory with the switch off, because the check that guards the extraction path was never added to this path. That is a bug on our side, not a design decision, and it is written here because a privacy control that is 90 percent effective while being described as total is worse than one honestly labelled. Until it is fixed: if you keep memory off, check the memory list in Settings from time to time and delete anything you did not want kept — deletion works normally, and every entry is listed and individually removable.
What it is told not to store#
The extraction prompt instructs the model to keep study context only, and never to store health, address, phone, financial or relationship information.
That is an instruction, not a filter
Nothing inspects the extracted text afterwards and rejects it. It is a strong default and it holds in practice, but it is a model following a rule rather than a guarantee enforced in code. Treat it that way: if there is something you do not want written down, keep it out of the chat, and check the memory list — every entry is visible and individually deletable.
Turning it off and deleting it#
- Settings, then Memory. The toggle Remember things about me stops the extraction path from saving anything new; existing entries stay until you remove them. See the warning above for the one path it does not yet cover.
- Every stored memory is listed under that toggle with its own delete control.
- Forget everything removes every stored memory at once.
- Deleting your account from Settings, then Account, erases the account and everything tied to it. That cannot be undone.
Where memory applies#
User memory is recalled into ordinary chats. Chats scoped to a workspace use that workspace's own knowledge base instead, so work you keep in one workspace does not leak into unrelated conversations.
Written from the code that implements it. Verify against src/lib/m/userMemory.js, src/components/m/settings/SettingsMemory.jsx.