The context that compounds

Private bankers figured this out decades ago. The best ones kept what they quietly called context cards — brief notes written after every client meeting, usually before they'd left the parking garage. Not summaries. Just fragments: mentioned her daughter's surgery next month, prefers not to discuss equities until Q4, asked about this fund twice now but always stops short of committing. Small. Specific. Forgettable if not written down. The discovery was empirical and unglamorous: clients who felt remembered generated twice the referrals, stayed twice as long, and trusted recommendations they'd have questioned from anyone else. Not because the banker was smarter. Because the banker knew things.

What those cards were doing is obvious in retrospect. They were capturing context — and context compounds.

What compounds, exactly

A single captured detail is nearly worthless in isolation. Your AI knowing that a colleague prefers structured agendas is, by itself, a mildly interesting fact. But put it next to three other things — she's under pressure on her own team right now, she's been short in her last two email replies, this conversation is about a deadline she doesn't control — and the briefing changes completely. The draft email changes. The timing changes. The ask changes.

None of those pieces would have registered as important when you captured them. The preference was mentioned offhand in a conversation weeks ago. The pressure was implicit in a Slack message you skimmed. The pattern in her email replies was something you noticed but didn't consciously log. But if you'd written each of these things somewhere they could be held together — and something was reading that place before you hit send — you'd be working with a fundamentally different kind of tool.

This is adjacent to what Daniel Levin and colleagues found in their 2011 research on dormant ties. Reconnecting with someone you haven't spoken to in years is genuinely valuable — often more so than reaching out to active contacts — but only when both parties can recover enough shared context to rebuild trust quickly. Reconnections fail most often not because of awkwardness or changed circumstances, but because of a specific failure: neither party can remember well enough where they stood. The shared history is there. The context to activate it isn't.

Robin Dunbar's research on cognitive limits suggests we can maintain active awareness of roughly 150 people at once. But most people who've spent a decade in any industry know far more than 150 people whose context actually matters. Those relationships don't disappear. They go dormant. What turns them back on is being able to pick up the thread. That requires context — and context requires somewhere to live.

The asymmetry you don't notice

Here's the counterintuitive part: the AI with less specific context gives you the better-sounding answer. Polished, balanced, confident. The AI with more specific context gives you a slightly rougher answer that happens to be exactly right.

"Draft a follow-up to Sarah about the timeline slip" asked of a context-free AI produces competent professional prose. Asked of an AI that knows Sarah, her communication preferences, and what she said in your last meeting, it produces something usable. These are different things, but they look similar until Sarah reads it.

This asymmetry is why it's so easy to underestimate the value of accumulated context. The payoff is invisible when the context doesn't exist. You never see the better email that didn't get written because you hadn't captured anything. You just see the generic one that sent fine and landed neutrally.

The capture problem

The reason people don't build up personal context stores is that the relevant thing never feels worth capturing in the moment. Sarah prefers not to be surprised in front of her team — you'd remember that, wouldn't you? Probably. For six months. Then you'd remember having learned it once, but not the specifics. Then it would be fully gone, leaving only the faint instinct to be careful around Sarah without quite knowing why.

This is the structural problem. The context that compounds is usually the context that felt too small to bother with. The thirty-second note after a call. The preference mentioned in passing. The reason you made a decision that seemed obvious at the time.

Writing these things down doesn't feel productive. It feels like overhead. But that's because we evaluate capture by how it feels to write, not by how it would feel to have — three months later, with a draft in front of you and a deadline looming.

What this changes

Business CRM systems exist because companies discovered empirically that keeping context creates durable advantage. A salesperson with good notes outperforms one without. Not because they're smarter or more personable. Because they can use the right fragment of context at the right moment, and that changes the interaction.

This never transferred to personal life, which is strange. Your doctor keeps notes about you. Your financial advisor keeps context. Your AI assistant — in most configurations — knows nothing. Every conversation starts over.

The bet behind Harbor is that an AI reading from a year's worth of your specific context — person records, preferences, relationship notes, decision rationale — produces qualitatively different results than the same model with no context at all. Not louder or faster. Just more accurate. More yours.

The private bankers figured this out in the parking garage after meetings. The insight still holds.

---

Asgeir Albretsen is the founder of Harbor.