The dashboard's AI is a Clippy — but it's a Clippy that reads four memory slots, and what you write there decides whether every downstream feature is brilliant or useless.
The memory architecture has four blocks. Business description: who you are, in prose (his was drafted by Claude from a web-scrape plus his service list). Business offers: what you sell, split into separate toggled entries with pricing, so the AI learns the difference between tiers instead of digesting one blob. Customer profile: the ICP — problems solved, benefits, USPs, goals, success stories, and the keyword include/exclude lists that steer lead search. Guidance: the behavioral rules — tone, format taboos, what a CTA looks like.
The payoff structure matters: the Copilot itself is weak at generation, but memory feeds everything else — super search returns ICP-matched leads, campaign ideas come out on-strategy, and the sales/reply agents answer prospects with correct pricing without being re-briefed. Memory is written once and consulted by every feature; that's why it's the first hour of the session, not an afterthought.
The live before/after: his year-old memory ('vague — I don't know what I was doing back then') versus the rebuilt version, which immediately powered a million-lead ICP-matched search.
This is context engineering with a form UI — the same signal→context→decision loop from Session 6, where memory is the context stage and every campaign is a decision made from it.
The Copilot is the product — prompt it well and it does the campaign.
The Copilot is a Clippy; the MEMORY is the product. Weak generation, strong context — write memory carefully and use real AI tools for the copy.
It's your own Clippy. It's Instantly's own Copilot — a very small note-taker. Nail the memory, though, and it knows everything about you.0:25:36
Go deeper
In one line: Instantly's Copilot is 'your own GPT inside the dashboard' — honestly rated as 'your own Clippy,' far dumber than a real Claude/ChatGPT project — but its MEMORY is the foundation of everything: business description, business offers, customer profile (ICP), and guidance. Nail the memory and every downstream feature (lead search, campaign ideas, the sales and reply agents) inherits full context of who you are and what you sell.
The Session 7 positioning work maps slot-for-slot into memory: the offer-construction output IS the business-offer entries (0:25:36)
Proof it works: after memory setup, 'find lead prospects' returned 1,000,000+ leads matched to his exact ICP with zero extra prompting (1:57:28)
Settings: keep analytics enabled; optional Slack connection for campaign updates; tasks can schedule recurring pulls like weekly analytics (0:27:37)
A saving gotcha bit him live: unsaved memory sections silently lost progress — save each section as you go (1:18:35)
▶ Watch this taught: 0:23:34
Answer from memory first — the recall attempt is what makes it stick. Then reveal.
Name the four memory blocks and the features they feed.
Business description, business offers, customer profile (ICP), guidance — feeding lead search, campaign ideas, analytics, and the sales/reply agents.
Why split offers into separate entries instead of one pasted block?
Instantly trains on the distinction between tiers — separate toggled offers with prices let agents quote and route correctly without extra context.
