Most AI assistants start fresh every conversation. You explain your situation, your team, your goals — again. She does not work that way. Every conversation, every task you complete, every meeting you create builds a richer picture of you. And she carries that picture into everything she does.
This page explains how that works — what she stores, how she retrieves it, and why the result feels less like a search engine and more like someone who has been paying attention all along.
Memory is not a single thing. It is three layers working together.
Human memory works in layers too — what you're thinking right now, what you know well, and what you've experienced over years. Her memory is designed the same way. Each layer serves a different purpose and a different timescale.
What you're focused on this conversation — your current goal, your mood, the last few things you said. Gives her instant awareness of where you are right now.
A living document of who you are — your identity, active projects, the people you work with, your preferences, recurring patterns, and key milestones. Consulted on every message.
A structured map of everything across all your conversations — people, projects, metrics, events, ideas — connected by relationships and anchored in time. The long memory.
Eight kinds of things she tracks about you.
The knowledge graph is not a notebook of quotes. It is a structured map of the world around you — people, projects, habits, events, metrics, ideas, places, and preferences — each understood as an entity with properties, and each connected to the others.
Co-founders, investors, advisors, clients — with roles, how you know them, and what's been said.
What you're building, its status, priority, and where it stands today versus a month ago.
MRR, DAU, runway, growth rate — tracked over time with their direction and context.
Pitch meetings, product launches, calls — with outcomes, participants, and follow-ups.
Things you've floated — features, pricing changes, strategic pivots — and their status.
Where you're based, where you travel, and the significance of each location to your work.
What you like, how you work best, things you've said you want more or less of.
Patterns in how you work — when you're sharpest, what you do weekly, how you prefer to communicate.
It is not enough to know that Sarah is an investor. She knows that Sarah led your seed round, that she knows your advisor Omar from YC, and that the last time you spoke was before your demo day. The relationships between facts are what make her useful — not the facts in isolation.
Every message teaches her something.
Two things happen in parallel when you send a message. Before she responds, the system searches your knowledge graph for the most relevant facts about you and injects them into her thinking. After she responds, the conversation is automatically analyzed and the new facts are written back into your graph. You never do either step yourself.
Just got off a call with Sarah. She wants to lead our seed at a $3M cap.
That's a strong signal from her — she doesn't usually move this fast. Does she know you're also talking to the Flat6 partners?
She recalled: Sarah is an investor you've been in discussions with. She knows your current fundraising context and who else is in the conversation.
She doesn't. I want to close Sarah first.
Makes sense. Do you want me to draft a term sheet response that holds the others off while you finalize with her?
After this exchange: Sarah's role (lead investor, $3M cap), the event (seed round call), and the outcome (positive, closing in progress) are written to your graph.
She knows what was true then, not just what is true now.
Most memory systems overwrite old information when something changes. If you tell her you moved from Dubai to New York, she forgets you were ever in Dubai. Her memory does not work that way. Every fact is anchored to when it became true and when it stopped being true. History is preserved.
If you ask her "where was I based when we raised our seed?" she can answer correctly — even if you've moved twice since then.
She learns from what you do, not just what you say.
Most AI assistants only know what you've told them in chat. She also watches your workspace. When you complete a task, create a meeting, add a contact, or your metrics update, that information flows into her memory automatically. You don't have to narrate your own work to her.
What you finished, when, and how it relates to your active projects.
Meetings created, participants, and what they're connected to.
Revenue, growth, customers — ingested automatically and tracked over time.
New people in your world, who they are and how they connect to your work.
None of this requires any action from you. There is no "save to memory" button, no tags to add, no summaries to write. The workspace listens and learns. You just work.
She does not pull everything for every message. She reads the room.
A simple "good morning" does not need a full briefing on your investor relationships. A question about your growth trajectory does. The system classifies each message before searching, so the depth and richness of memory retrieved matches what the question actually needs. Responses stay fast and focused.
She doesn't match keywords. She follows relationships.
The difference between a good memory and a great one is the ability to connect things that were never mentioned together. Because her memory is a graph of relationships, she can reason across chains of facts — not just recall individual ones.
You ask: "Should I reach out to anyone about our Series A?" — She can reason from your MRR trend, to the investors she knows are active at your stage, to the warm intros available through the people already in your graph, to the specific timing of your last conversations with each of them. That is not keyword matching. That is reasoning across your actual network.
She gets better every day you use her. That is the point.
Memory is not a feature. It is the difference between an assistant that feels like a tool and one that feels like a partner. Every conversation, task, and decision you work through with her makes her more useful to you — not because she was programmed with your life, but because she has been paying attention to it.
1 Memory is stored encrypted at rest. The knowledge graph is subject to the same zero-knowledge guarantees as the rest of your workspace. See How the vault works.
2 You can clear your memory at any time from account settings. Clearing deletes your knowledge graph, your profile document, and all working memory. It cannot be undone.
3 You can also pin specific facts — things you always want her to know, that will never decay or be overwritten by a consolidation pass.
4 Memory can be disabled entirely from settings. When off, she operates on the current conversation only, with no long-term recall.