How your assistant knows you before it knows anything

I opened a fresh chat the other day with an assistant that remembers me, and it greeted me already knowing things. My projects, the thread I left open, what I was supposed to be working on. One turn in and it was oriented, no pause and no questions.

It’s a good feeling. It’s also worth asking where it came from, because someone made that choice for me, and it’s quietly the most important choice in the whole system.

The first move nobody talks about

Every assistant that remembers has to answer the same question at the start of a session: how does it know anything about me before I’ve said a word? When I built it both ways, the field split in two.

The first is to push. Before I arrive, the system writes a summary of me and drops it in the assistant’s lap, so it opens already holding a little dossier. That’s the warm start you feel. It knows you because someone handed it a note.

The second is to pull. The assistant starts empty and reaches for what it needs, asking a memory store the moment something I say calls for it. No note waiting on the desk. It begins knowing nothing and goes to look.

Same goal, opposite stance: one arrives knowing, the other arrives ready to find out.

What the warm start actually costs

The pushed summary feels better at first contact, and that feeling is the problem.

A summary is written ahead of time, so it’s already a little stale, already guessing what will matter this session. It’s also lossy by design. It can’t carry everything, so it carries a compressed gesture at everything, with the reasoning sanded off and the links between things gone. The part that bites is that the assistant can’t tell. It reads the summary as knowledge, answers fluently from it, and sounds completely sure. You get confidence built on a paraphrase of a paraphrase.

So the assistant that opens already knowing you is also the one most likely to be confidently wrong about you, because it never had to check. It was handed a story, and it trusts the story.

What the cold start buys

The other stance has an honest flaw, and I’ll put it on the table first.

A memory store can’t tell the assistant to check it. The knowledge sits there, but it can’t reach out and announce itself; something has to decide to go and look. So a pulled start needs a nudge, a small standing instruction that says reach for this when it comes up, and that nudge has to come from outside the store. It’s a real seam, and pretending otherwise would be dishonest.

Once the nudge is in place, look at what you get. Everything the assistant knows, it knows because it went and looked, just now, in answer to what I actually said. A fresh pull, the reasoning still attached and the connections intact. It starts colder and catches up in one move, and what it catches up to is the real record rather than a summary’s silhouette of it.

The bet worth making

I’d rather my assistant start knowing nothing and look than start knowing a summary and believe it.

The warm start optimizes for the first second of the conversation, and being recognized that fast feels good. But I’m not in this for the first second. I’m in it for the part where the assistant reasons about my actual situation, and there the pushed summary keeps charging a tax, every session, in small confident errors I have to catch. The cold start pays its cost once, at the seam, and stays honest the rest of the way. It earns what it knows instead of being handed a guess. That only looks like the worse deal if you judge memory by how it feels on the opening line instead of whether it holds up when it counts.

So the next time something greets you already knowing you, the question isn’t how much it remembers. It’s whether it went and looked, or whether it’s just gotten very good at sounding like it did.