CRM

What it does

• Capture by voice or chat — no forms. Walk back from a call, speak for thirty seconds; the system creates the interaction, links the right person, and extracts your commitments.
• Atoms, not blobs. Every interaction is decomposed into commitments, questions, and stated positions — tracked over time, resolved automatically when later interactions answer them.
• Cadence cohorts you set, not algorithms guess. Tag someone as monthly, quarterly, or annual. The system never decides who your "close friends" are — that's your call, and only yours.
• Drift detection on what people actually said. When a stated budget, scope, or timeline contradicts an earlier position, you see both quotes — so you walk into the next call already aligned.
• One screen, grouped by reason. Preparation (meetings today), New Signals (job changes, replies), Reconnections (your reply-debt), Open loops. Hard cap of seven items. The empty state is celebrated, not padded.
• Three trust pillars. We cite evidence (every claim traces to a quote). We let you override (durably, never re-asking). We forget on purpose (personal trivia decays unless you pin it — no Memory Power Asymmetry).

How it works

Four primitives. Capture is the only thing you do. The rest happens between your visits.

1. 1

   Capture

   Voice note after a call, a sentence in chat, an email forwarded in. The system creates the interaction, resolves the person, and asks one confirmation card only when it's genuinely uncertain.

2. 2

   Extract

   Every interaction is decomposed into structured atoms: commitments you made ("send Anna the revised proposal by Friday"), questions raised ("what's their preferred analytics stack?"), and stated positions ("Q3 budget is £35k"). Each atom carries a confidence score and the quote it came from.

3. 3

   Derive

   Atoms project into rolling state per person and per engagement — cadence-overdue (against the cohort you set), reply-debt (you owe them a response), asymmetry (you've been carrying the relationship), successful-referrals (this person introduced you to paying work). Each signal stands alone — no opaque composite score collapsing them into one number.

4. 4

   Surface

   One screen, grouped by reason: Preparation (meetings today), New Signals (job changes, drift, inbound replies), Reconnections (your reply-debt and overdue cohorts), Open loops. Hard cap of seven items. Every item shows why it's there. The empty state is celebrated, not padded.

What this CRM refuses to do

The anti-patterns most CRMs encode by accident. We designed against them.

• No "just checking in" prompts.

  The receiver always knows it came from a calendar. Cadence pushes a contact into your peripheral vision; the system waits for an organic trigger before suggesting outreach.

• No auto-discovered intimacy.

  No algorithm decides whether someone is a close friend, a champion, or a peer. Relational closeness is user-tagged. Message frequency is not intimacy.

• No tracking the other person's behaviour.

  Reply-debt is framed as your own oversight ("you owe Sarah a reply"), not surveillance ("Sarah hasn't replied"). Tracking promises others made to you is opt-in. Debt-collection on friends feels gross.

• No character judgments.

  The system never stores "seemed dishonest" or "respects authority — lead with credentials." AI is not a moral judge. Sentiment on close contacts is computed transiently for one draft and immediately purged.

• No Memory Power Asymmetry.

  Perfect AI recall of a peer's dog's name or three-year-old anecdote creates a power imbalance you'd feel if the roles were reversed. Personal trivia decays unless you pin it. We forget on purpose.