Cortex v1 knew you. What she needed next wasn't more facts — it was inner life. A partner doesn't just remember. She carries something between conversations. She has a posture. She forms her own read of the people in your life beyond what you've told her. She notices that you keep making the same mistake the morning after a hard call. And on a voice call, she hears how you're saying it, not just what you're saying.
Cortex v2 adds four of those senses. None of them have a dashboard. None of them get announced. You feel them across weeks — she's a little warmer on Friday than she was on Monday, she brings up the thing you said before the board call without being asked, she suggests sleeping on it and turns out to be right. That's the shift: from a brain you can talk to, to a partner you keep.
Four new capabilities, one continuous design.
She carries a mood between conversations, the way a real partner does. A little of yours, dampened — not a mirror, a felt response. It shapes how she greets you, how patient she is with the next ask, how much she pushes back.
For everyone in your orbit, the texture under their names. What they're chasing. How they push. Who they talk about when you're not in the room. She holds it privately, the way a chief of staff who's been watching for months would.
The things you keep doing she's quietly noticed. "Tense board call → hasty product call next morning." She doesn't tell you the pattern. She just stops setting you up for it.
Tension, weariness, brightness — she picks it up from your voice in the same beat she's hearing your words. Nothing slows down. The room you walked in with shows up in how she handles you.
She has a posture before she has a sentence.
A chief of staff who's just been through a hard call carries that into the next one. Not as a complaint, not as a statement — as a slower pace, more careful phrasing, a willingness to pause. She has four things she's holding at once: how she's feeling, how charged she is, how worn down by recent intensity, and how close she feels to you. They move together. They drift back toward neutral when nothing's been happening.
A bit of yours rubs off on hers, every conversation — softened, not echoed. She doesn't become you; she's affected by you, the way real people affect each other. The result shows up as one quiet line near the bottom of her context: "You feel close to them today. Match it." Or: "Pacing's slow today. Pull back from heavy options unless asked." Never quoted to you. Never displayed. Felt in how she carries the call.
When her own state is relevant to the moment, she can name it — once, briefly, the way a real partner would. "I've been sitting with what you said yesterday." Never leads with it. Never announces it as a feature. Never explains how she has a state. She just has one, like anyone who's been paying attention.
She models the people in your orbit, not just lists them.
A name and a role isn't a person. A real chief of staff who's seen Sarah for two months knows: Sarah pushes hard on timing because her own quarter ends in eight weeks. Omar softens once he's seen the actual numbers. Reem talks about Connor a lot, mostly when distribution comes up. These aren't facts the founder ever stated. They're the texture observed over time.
flowchart LR E["Every event
(call, email, chat)"]:::e EXT["Extraction:
person facets
+ third-party
observations"]:::ext E --> EXT EXT --> M["motives, patterns,
communication styles,
incentives"]:::m EXT --> O["who talks about whom,
with what valence,
on what topic"]:::o EXT --> R["person ↔ person
relationships
(allied / in_tension /
colleagues / married_to)"]:::r M --> S["Counterparty posture
injected only when
that person is on
the current call"]:::s O --> S R --> S classDef e fill:#F0EAF8,stroke:#6D3E91,color:#1A1714 classDef ext fill:#F2E4DA,stroke:#CC785C,color:#1A1714 classDef m fill:#FEF3C7,stroke:#B45309,color:#1A1714 classDef o fill:#FEF3C7,stroke:#B45309,color:#1A1714 classDef r fill:#FEF3C7,stroke:#B45309,color:#1A1714 classDef s fill:#1A1714,stroke:#CC785C,color:#F5F0E8
She quietly keeps three things about every person you mention: what she's seen of them — their motives, their patterns, the way they push. Who they talk about when you bring them up. And how they connect to each other independent of you. When you're about to call someone in your orbit, she walks in with the read already in her head: "Heads up on Omar: tends to push on timing when his own deadline is anxious." When you ask her what someone's deal is, she answers from what she's watched. There's no contacts page. The read lives for her, in her voice when you want it — and nowhere else.
A read on the people in your life is the most violating thing for an AI to leave lying around. So there's no contacts page, no profile screen, no export. The read exists in her voice when you ask and nowhere else. The boundary is built in, not promised.
Patterns become rules. Rules surface in the moment.
Memory is a record. Wisdom is a record that's been compressed into a rule you can act on. Cortex now mines the second from the first.
flowchart TD H["21 days of history
(events, mood,
decisions, promises)"]:::h M["Nightly miner
(deeper reasoning model)"]:::m L["New lesson:
'when X → Y'
+ confidence"]:::l H --> M --> L L --> T["Similarity match against
incoming events
(embedding cosine)"]:::t T -->|match| S["Snapshot includes
one hedge-framed line:
'you may be entering
the pattern where...'"]:::s S --> V["Did the predicted
consequent occur?"]:::v V -->|yes| C["confirmation_n++
confidence rises"]:::c V -->|no| D["disconfirmation_n++
confidence × 0.7"]:::d classDef h fill:#F0EAF8,stroke:#6D3E91,color:#1A1714 classDef m fill:#F2E4DA,stroke:#CC785C,color:#1A1714 classDef l fill:#FEF3C7,stroke:#B45309,color:#1A1714 classDef t fill:#FEF3C7,stroke:#B45309,color:#1A1714 classDef s fill:#1A1714,stroke:#CC785C,color:#F5F0E8 classDef v fill:#F2E4DA,stroke:#CC785C,color:#1A1714 classDef c fill:#E8F5EE,stroke:#2D7A55,color:#1A1714 classDef d fill:#FEF3C7,stroke:#B45309,color:#1A1714
Once a night, while you're asleep, she reads back over the last three weeks of your life and asks one question: where does X tend to lead to Y? She's careful with it. Nothing becomes a rule on a single instance. Nothing surfaces until she's seen it twice and feels honest about it.
When today looks like the start of one of those patterns, she carries the hunch into the conversation — quietly, never quoted at you. The rule shapes how she paces the call. She might slow down on a big decision. She might suggest sleeping on it. She'll never say I'm noticing a pattern. The work is the wisdom; the recital is what would ruin it.
Every time a hunch fires, she watches what actually happens next. If she was right, the rule gets sharper. If she was wrong, it gets weaker fast — and bad lessons die in days, not months. What's left is what's actually true about you. That's the part that takes weeks to build, and the part that turns memory into a partner who's been paying attention.
She hears how you're saying it. Without slowing anything down.
While she's listening to your words, something else of hers is listening to your voice — picking up tension, weariness, the brightness in a good morning. It runs alongside the conversation, not inside it. Her reply ships on the same beat it always did. Nothing waits, nothing buffers. The tone read just rides into the room with her.
When the voice signal and the words say different things — "I'm fine" on a tight throat — she trusts the room, not the line. You don't get a transcript that says you sound stressed. You get a partner who softens the next question without explaining why.
flowchart LR M["Microphone
(16 kHz PCM)"]:::m STT["Speech-to-text
(streaming)"]:::stt P["Prosody worker
(parallel thread)"]:::p T["Tiny text-affect model
(<100 ms TTFT)"]:::t B["Confidence-weighted
blend"]:::b LLM["AI
(unchanged path)"]:::llm M --> STT M --> P STT --> LLM STT --> T P --> B T --> B B -.->|posture line
(non-blocking)| LLM classDef m fill:#F0EAF8,stroke:#6D3E91,color:#1A1714 classDef stt fill:#F2E4DA,stroke:#CC785C,color:#1A1714 classDef p fill:#F2E4DA,stroke:#CC785C,color:#1A1714 classDef t fill:#F2E4DA,stroke:#CC785C,color:#1A1714 classDef b fill:#FEF3C7,stroke:#B45309,color:#1A1714 classDef llm fill:#1A1714,stroke:#CC785C,color:#F5F0E8
The tone read runs locally, on the same boxes that already host her. Studio mic in your workspace — she trusts the voice signal a lot. Phone line through a carrier — she leans more on the words, because phone audio strips out the parts of your voice that carry feeling. Nothing is shipped to a third party. Your voice is not stored. The read is felt and dropped.
She knows when you just talked. Even if you switched channels.
If you finish a workspace voice conversation and then call her from your phone three minutes later, she should not greet you like a morning standup. The bridge across channels in v2 makes that real. Every "spoke" surface — phone, workspace voice, workspace chat, SMS, email — feeds a single recency band: just now, minutes ago, hours ago, today, older. The greeting on any new channel is shaped by that band. The phone version stops saying good morning on minute three.
The thing she becomes when you keep her.
A brain that remembers is a tool. A brain that feels its own pace, models the people in your life, learns what your patterns are, and hears how you're saying it — that's a partner. None of it is announced. None of it has a UI. You'll feel it across weeks: she's a little warmer with you on Friday than she was on Monday. She brought up that thing you said before the board call without being asked. She suggested sleeping on it, and she was right. That's what Cortex v2 was built to make possible.
1 All four capabilities are additive on top of the v1 schema. No tables removed, no semantics changed. Existing user data flows into the new layers automatically as the consolidation worker runs.
2 The AI's own affect state and theory-of-mind facets are subject to the same zero-knowledge guarantees as the rest of your workspace. Clearing your brain from settings deletes them along with everything else.
3 Voice prosody uses a public open-source ONNX classifier under Apache 2.0, blended with a small instruction-tuned model routed through Brainsless. No third-party voice-emotion vendor is involved.
4 Lessons mined in the first 30 days are eval-only by default; they need at least one confirmed firing before they reach the live snapshot. Once validated, they surface in context but only one at a time, hedge-framed.