The contact brief
Before Connect answers, it is handed a short block about the person: their name and how they prefer to be addressed, their company and city, their stage, their language, a few facts, the tail of the notes on file, and up to three standing instructions. It is budgeted at around 600 characters, because on a live call every extra character is time the caller spends waiting.
What the block says#
- *Their name is …*, with *(they prefer to be called …)* when a preferred form is on file.
- *They are from …*, with the city when it is known.
- *Relationship stage: …*, so an answer to a long-standing customer does not read like an answer to a stranger.
- *They prefer to speak …*, when a language preference has been recorded.
- Facts on file, one per line, each value cut at 160 characters.
- *Notes on file:* the last 400 characters of the notes, passed through the name guard.
- Up to three standing instructions for this person, read from memory as directives.
When none of that exists the block says so plainly — *nothing on file beyond the number* — rather than being empty. A model given an empty section will invent one; a model told the section is empty does not.
Budgets, and why they are small#
| Block | Budget |
|---|---|
| The contact block | about 600 characters |
| Memory brought into an answer | about 700 characters |
| Knowledge brought into an answer | about 2,000 characters, and about 4 facts |
| A human voice profile | about 2,500 characters |
| Notes tail inside the contact block | the last 400 characters |
| One fact value | 160 characters |
Those are not arbitrary. Instruction size and the time to the first spoken word move together: a briefing of about 7,500 characters measured 1.2–1.8 seconds to first token, and one of about 9,600 characters measured 2.3–3.4 seconds. On a phone call that difference is the gap between a natural pause and a silence the caller starts talking into.
What is deliberately kept out#
- Connect's own markers. The test-call flag was once reaching the model as a fact line and being re-billed every turn. Internal markers are filtered by name now.
- A form of address mistaken for a name. Facts keyed as a name whose value is an honorific are dropped by the same guard that rewrites summaries.
- Empty values. A fact with no content is not rendered as a blank line.
- More than three standing instructions. Directives are ranked and cut; the block is a briefing, not the whole memory.
The general rule is that everything in the block is something an answer might reasonably use in the next sixty seconds. Anything that is only interesting later belongs in memory or on the timeline, where a person can read it without a model paying for it on every turn.
When the brief is wrong#
Check the person the call resolved to.
Result A brief about the wrong human is an identity problem, not a briefing problem — Connect matched the wrong person covers it.
Read the contact tier of memory for that person.
Result Standing instructions in the block come from directives; removing one there removes it from the next briefing.
Read the notes on file.
Result Only the tail is used. A correction added at the end reaches the block; one buried thousands of characters up does not.
Check for a block on the channel.
Result Blocked channels are resolved alongside the brief, from memory tags rather than from a field, so a block set in the right place holds across future conversations.
Questions#
Why is only the tail of the notes used?
Because the briefing has a budget and the newest writing is the likeliest to matter. If something must be visible to every answer, put it in memory at the contact tier rather than at the top of a long note where it will fall outside the window.
Does a longer brief make answers better?
Not reliably, and it always makes them later. Measured first-token time rises with instruction size, and on a call that shows up as dead air. The budgets exist so the useful lines are not competing with lines nobody reads.
Is the same block used on email and WhatsApp?
The idea is shared — a short, budgeted description of the person, assembled from the same records — but the tightest budgets belong to the voice path, where latency is heard. What is in a written reply is grounded in the same memory, knowledge and relationship records.