vMira now remembers what matters across conversations — the projects you are working on, the tone you prefer, what it has already explained to you. Memory is on by default for paid plans and off by default for the free tier. Whatever the default, it is yours: viewable, editable, exportable, and deletable, in one screen.
Episodic memory works like a short notebook. After every conversation, the model proposes a small number of notes — facts, preferences, recurring projects — and writes them with a 200-character ceiling per note. You can review, edit, mute per-conversation, or delete each note from a single page in your profile. Nothing is buried; nothing is encrypted in a way that hides the contents from you. The notebook is the source of truth for what the model remembers about you, and you can read every line of it.
What it stores and what it does not
The model is trained to store three kinds of facts. Your work: the languages you write in, the tools you use, the deadlines you have mentioned, the people on your team. Preferences: how you like explanations (short or detailed, with or without code), what tone to avoid, what topics you do not want unsolicited advice on. Recurring context: projects in flight, books you are reading, problems you are turning over. It is trained explicitly not to store: secrets (credentials, keys, passwords) even if you paste them; sensitive personal data of third parties (a colleague's phone number, a family member's medical detail) even if you mention them; and anything you flag as private during the conversation with a single command. The mute and delete actions are fully reversible — until you confirm — so accidental deletion is not a one-way door.
“A good memory is not a big memory. It is one that forgets the right things.”
Total control
Memory can be turned off entirely with a single switch in settings, in which case the model behaves identically to the pre-memory product. Temporary mode runs a single conversation without writing anything to the notebook — useful for sensitive sessions you do not want to leave a trace of. Export and import are first-class: download your notebook as JSON, edit it, and import it back into another account or a fresh install. Enterprise customers get an additional layer: per-team policies that restrict what categories of facts the model can store, organisation-wide notebook export for audit, and SCIM integration so that off-boarding a team member also retires their memory store automatically.
How the model decides what to save
After every conversation, a separate model pass — not the one that wrote the answers — reviews the transcript and proposes notes. The proposer is trained on a dataset of conversations paired with high-quality notes, where high-quality means useful in the next conversation, not just descriptive of the last one. The proposer scores each candidate note on three dimensions: durability (will this still matter in a month?), specificity (is this about you or about the world?), and necessity (would the next conversation be measurably worse without it?). Notes below the threshold are dropped silently. Notes above the threshold are written to the notebook and shown in your notes view, where you can edit or delete them. The proposer's recall is conservative on purpose: a missing note is recoverable in the next conversation, a stored secret is not.
Privacy filtering: why secrets never enter the notebook
The proposer is trained explicitly not to save certain categories of content even when they appear repeatedly. Credentials, keys, passwords, and authentication codes are filtered by a deterministic classifier that runs before the proposer sees the transcript. Sensitive personal data of third parties — a colleague's phone number, a family member's medical detail — is filtered by a learned classifier with high recall and a tunable threshold. Anything you flag as private during the conversation with a single command is hard-excluded from the proposer's view. The combined filter is itself a regression-tested category in our public evaluation framework, so it is checked on every model release. If the filter starts under-blocking, we do not ship.
How memory composes with reasoning
When the reasoning model is invoked in a conversation, the notebook is part of the context the model deliberates over. That has two effects. First, the model can refer back to earlier facts during a long thinking trace without re-asking — a multi-step plan can incorporate the user's stated preferences without prompting. Second, the model can add provisional facts to its working memory during a single reasoning trace without committing them to the notebook; if the conversation ends without confirmation, the working memory evaporates. This separation between persistent notebook and ephemeral working memory is what makes reasoning over personal context safe: nothing the model invents in a thinking trace becomes a permanent note unless the proposer agrees on the next-conversation pass.
Per-team enterprise policies
Teams plans add an admin console where the team owner can restrict the categories of facts the model is allowed to store. The categories are the same three the proposer learns from — work, preferences, recurring context — and each can be toggled team-wide. A common policy in regulated industries is to restrict storage to preferences only, so the model remembers tone and format choices but never commercial or operational facts. Notebook export at the team level is an admin-controlled action; it produces a JSON file with one notebook per user, ready for retention or deletion. SCIM integration ties off-boarding to notebook deletion: when a team member's account is deactivated, the notebook is purged after a configurable retention window.
Notebook export format
The notebook is a JSON document. Each note has an ID, a category, the text (≤200 characters), the conversation ID where it originated, the timestamp it was written, the timestamp it was last modified, and a soft-delete flag. The format is stable; we will accept old notebooks for at least two model versions after a format change. Importing a notebook into a new account or a fresh install is a single call from the settings UI; the model behaves immediately as if the notes had been generated in the importing account. For users who maintain multiple accounts (personal and work, for example), the export-import flow is the supported way to move context between them — there is no automatic syncing across accounts on purpose.
What memory is *not*
Memory is not a database for the model to query. The notes are part of the prompt context, not a vector store with retrieval. That is a deliberate trade. A retrieval-backed memory scales further but introduces a second model — the retriever — whose decisions about what is relevant are opaque to you. A prompt-context memory has a hard size limit (a few hundred notes for most users, thirty on the Free plan) but every note is visible at every step. We picked visibility over scale. Customers with genuine large-scale memory needs — operations teams that want to remember thousands of customer interactions — should use the API and build a retrieval system on top, where the trade-offs are theirs to control.