Awareness

Olyvr keeps three kinds of context about you, all stored on your machine and all in your hands. Open Awareness under Olyvr Mind in the sidebar (the page is headed Knowledge): facts about you in the notes, what your chats have taught it as learned preferences, and the rules you set as preferences. Olyvr reads all of it before answering: in every chat, for every Reflection card, when Improve rewrites a card’s prompt, and when it drafts the reply on a My Day item’s Reply card.

Notes

Five notes give the agent standing context:

Note	What goes in it
Me	Your role, team, and how you work
People	Colleagues, clients and their relationships to you
Projects	What you’re working on
Known aliases	Who you are in each system: work and personal addresses, GitHub login, Slack handle and name, Notion, Jira
Always	Instructions for every answer, such as tone, formatting or things to avoid

Each note holds up to 12,000 characters.

Write them yourself, or generate them

Type directly into any note. The lines you write are yours: updating from your data never corrects, merges or rewords them, while the rest of the note keeps growing and getting corrected as before. In Known aliases, what you add stays and is searched for too, and what you remove stays out.
Update from my data grows the Me, People and Projects notes from your workspace documents and from what your chats taught it about you, the people you mention and your projects. The notes are never rewritten: the model is given the notes as they are, line by line, and only the documents they have not read yet, and may only add a line (for a fact the notes do not hold) or correct one the data shows to be wrong or out of date. Olyvr applies those changes itself, so a line is never removed, and a correction that turns a line into something else is refused. In People and Projects each person or project has one entry, a heading with its lines under it: a fact the model adds about someone already there goes under their entry, under any spelling of the name, and the lines of each entry are kept together. Not everything is read, and not in arrival order: documents that mention you (by the names, addresses and handles your sign-ins and Known aliases give) come first, then your mail, chats and calendars, then your documents, newest first. Documents from Olyvr TeamMind, and documents older than a year, are read only when they mention you; the rest is left out, and the page says how many. One click reads up to twenty batches, three at a time; the rest is read in the background, a few batches a minute once no import is running, up to 2,000 documents a day. The page says when the notes were last updated, by which model, and whether there is more to read. Each document is read once, however often it is synced, and again if it changes. This needs a model and at least one imported document or summarized chat.
Known aliases is filled in at the same time, by search rather than drafting. Starting from your sign-ins, Olyvr searches your data (and any Olyvr TeamMind source that is connected) for each address, login and name it knows, takes the ones that appear next to a known one, and searches again with what it found, up to three rounds. In mail, only an address a message came from can be yours, never one it went to; a sender shown with your own name is yours, and anything less certain the model judges, listed only when a line of your data backs it. Like the Me note, these addresses go into every request. Unlike the other notes, this one keeps up on its own: once data has come in since it was built, and no import is running, Olyvr searches again from your sign-ins and everything it found before, so the note accumulates. Remove a wrong line and it stays out; Update from my data starts the search over from the sign-ins.

When your sources change, the notes say how many documents are still to read, and read them on their own. They go into every answer as they are meanwhile, the way learned preferences do, and your own text is always used. If documents arrive while an update is running (an import can take a while), they are simply read with the next batch; Known aliases is written from the data as it was and catches up on its own once the import is over. Only documents removed while a batch is with the model stop that batch, since the notes would describe what you took out.

Every version is kept

Before the Me, People or Projects note changes, by an update from your data, by a finished chat adding its facts or by your edit, the version going away is kept on this computer. Each note keeps its last 200 versions, which covers a day of background updates with room to spare. Turn it off in Settings → General → Keep Knowledge note history.

Learned preferences

Between the notes and the preferences, Learned preferences is a summary of all the summaries: meta-learning about you, written by your model, not by you. One tile says what your finished chats, read together with your notes, show about what you keep working on, asking for and deciding, and how you like it done. Under it is a list of the single patterns that summary rests on, grouped by dimension: communication, detail level, decisions, domains and expertise, working rhythm, recurring concerns, tools and formats.

Update now reads the summaries of all your finished chats and your four notes, and asks your model for the summary and the patterns. A pattern is kept only when its evidence is a line of a summary or a note; the rest are dropped. This needs a model and at least one summarized chat.
Hints, never orders. The summary and the patterns that are on ride along with every request, marked as observations. What you ask for in the request always wins.
Switch off any pattern that is wrong. It stays off through every rebuild, and the same pattern is not added again, even when the model words it differently. Forget one with the × beside it.
Correct it. Edit the summary and your version is kept until you empty it, when the next rebuild writes it again.
When it goes stale: each finished chat counts once towards the next rebuild, however often its summary is updated. After three the summary is marked as having new chats to learn from. It still rides along until you rebuild.

Preferences

Preferences are short rules, like “Keep emails under 150 words” or “Use British spelling”.

Each preference can apply to answers, documents, spreadsheets, decks, emails or everything.
Add an optional when phrase to apply it only in context. For example, “Use the formal template” applies only when your message mentions “board report”.
Turn preferences on or off without deleting them.

Update now reads your messages from your recent chats and suggests preferences. It only keeps rules it can back with something you actually wrote. It needs a model.

From your chats

When a chat is finished, Olyvr writes a short summary of it: what it was about, what was decided and what is still open. What the chat said about you, the people you mention and your projects goes straight into the Me, People and Projects notes, as a line each after what is already there. Facts already in a note aren’t added twice, and nothing you wrote yourself is changed. When you later ask about something related, the summaries of those earlier chats quietly inform the answer: they help the assistant understand your context, but it never quotes them, cites them or tells you it remembers an earlier chat. When you want to go back to a past conversation yourself, ask for it and it looks the conversation up.

When it happens: once a chat has been quiet for 10 minutes, or as soon as you move to another chat. Chats still unsummarized when you quit are picked up the next time Olyvr starts. Very short chats, such as a greeting, are skipped.
When you come back: if you reopen a chat and keep going, the summary is brought up to date once the chat pauses again. Olyvr only reads the new messages, together with the previous summary.
In a chat: the Knowledge button in the chat’s header shows what Olyvr remembers of it. Choose Remember it now to summarize it right away, or Don’t remember this chat to forget its summary and keep it out from then on.
Turning it off: Settings → General → Remember finished chats. Off, finished chats are no longer remembered and earlier chats’ summaries stop informing your requests. Facts already added to your notes stay there until you edit them.
A full note: each note holds about 12,000 characters. When a note fills up, the oldest lines that came from chats make room, never what you wrote. An update from your data adds nothing to a full note; instead the model may merge two or three lines about the same subject into one line that keeps every fact of each, which makes room a little at a time, and the page says the note is full. A merge is kept in the history like any other change, so the version before it can be put back.

Summaries are made by your model and stay on your computer. Deleting a chat deletes its summary.

From your Snapshot cards

Before a Reflection card’s prompt or answer is replaced, by a refresh, an edit, a saved Self Optimize rewrite or deleting the card, the version going away is kept: the prompt, the answer and when it was answered. It doesn’t go into your notes. It is there for the card’s History, for Self Optimize to learn from, and for the assistant to look up when you ask about what a card showed before. Each card keeps its last 30 versions. Turn it off in Settings → General → Keep Snapshot card history.

Which model

The Me, People and Projects notes are written by the best model your sign-ins serve, and by that model alone: a Fable or a GPT sol over an Opus, an Opus over a Sonnet (between providers, the one you signed in with first), with no fallback to another model. The page names it. If it fails, the page says so and the update is tried again later; nothing is written by a lesser model meanwhile. Set OLYVR_KNOWLEDGE_MODEL to pin another. Chat summaries, Known aliases, learned preferences and preferences use the summary chain in Settings → Model access, which spreads this background work over smaller, faster models.