Tell it once: memory and web search
By The Assistly team ·
There are two ways an assistant wastes your time on a live call, and they look nothing alike.
The first is repetition. You tell it to keep answers short. The next answer is short. The one after that, in tomorrow's call, is four paragraphs with bullet points, and you type "shorter" again. You are not getting an assistant, you are getting a stranger who is briefly agreeable.
The second is confidence about things it cannot know. You ask what a competitor charges and get a number from whenever the model last read the internet — stated in the same steady voice as everything else, with nothing to tell you it is eighteen months stale.
Memory fixes the first. Web search fixes the second. Both are on now, and neither asks you to set anything up.
Say it once and it sticks
Say what you want and Assistly writes it down.
you · i like shorter answers
Got it.
⌗ Remembered in Sales calls · "Keep answers short" · Undo
No command, no settings page, no magic word. "Stop using bullet points", "always show me the code first", "we're per-seat, billed annually" — if you'd expect a new colleague to remember it after you said it once, Assistly does too.
That last part took some getting wrong to get right. The first version only saved when you used the word remember, which meant the sentence people actually type — "i like shorter answers" — got a polite "got it" and changed nothing. Saying it is the instruction. The word was never the point.
Two kinds of memory, because they do different jobs
A memory is either a fact or an answer style, and Assistly files each one as it saves it.
- A fact is something true about you or your work: Pricing is per-seat, billed annually. I'm the CTO, not the founder. It sits in the background and informs answers when it's relevant.
- An answer style is a rule about the writing itself: Keep answers to three lines. Never use bullet points. It isn't background — it's an instruction that overrides how Assistly would otherwise write.
The distinction is the whole reason memory works. A rule filed as a fact is read as trivia and politely ignored, which is what "it says it's using my memory and the answers are still enormous" feels like from the outside. Filed as a rule, it wins.
New instructions replace old ones
Change your mind and the old rule goes, rather than sitting next to the new one:
you · actually, i like longer responses
⌗ Remembered in Sales calls · "Keep answers longer" · replacing
"Keep answers short"· Undo
An assistant holding two contradictory rules doesn't split the difference — it stops and asks you which one it should be following, in the middle of a call. So it doesn't hold two. And the swap is named on screen, because a rule that disappears quietly is worse than one that appears quietly.
And the thing you just said always wins
Memory is not a cage. Ask for more detail on a specific answer and you get more detail, standing rule or not — then things go back to normal on the next question. What you're saying now beats what you said last month, every time.
Where memory lives, and how to look at it
Memory belongs to a mode. Your interview mode remembers you're targeting staff-level backend roles; your sales mode remembers your pricing and that you never discount past 15%. Neither leaks into the other, because "keep it short" in a sales call and "keep it short" in a technical interview are not the same request.
Two consequences worth knowing:
- Memories apply while their mode is active. A memory saved into a mode you aren't using is doing nothing — the same rule that already governs a mode's reference documents.
- Every answer shaped by memory says so, with a Used 3 memories badge above it. An answer quietly steered by things you can't see is the thing worth being suspicious of, so we made it visible.
Open Modes → Memory and you get the list: every sentence, verbatim, in the order Assistly reads it. Reword one, forget one, forget all of them, or add one by hand. Entries Assistly saved on its own are marked, so what you told it and what it decided to keep are never confused.
That list is the honest answer to "what does it know about me?" — not a summary of what it knows, the actual text it reads.
When it isn't sure, it asks
Not everything people type is an instruction. "Shorten this one" is about the answer in front of you. "The call is at 3pm" is true today and false on Monday. Neither belongs in memory, and Assistly saves neither.
Between those and a plain instruction there's a middle: something you implied rather than stated. There, it asks instead of deciding — a small card under the answer with the exact sentence it would save, and two buttons. Say no and it won't ask about that again.
The card only appears as a reaction to an answer, never on the first thing you type in a thread. Opening a conversation with "make it shorter" is a brief for the task at hand; saying it after reading something is how people actually express a preference.
Web search, built in
The other half. A model's training is frozen on a date, and a great deal of what gets asked on a live call is newer than that: what a competitor charges now, what shipped in the latest release, who runs a company this quarter, whether a rumour is true.
Assistly searches the web itself when the answer depends on any of that — during the answer, not as a separate step you have to ask for. The answer cites what it used, inline, and carries a Searched the web badge so you know where it came from.
Nothing to connect and no key to paste. It's there on every account, on typed questions, on Assist, and on questions it picks up from the call.
It also knows when not to search
This matters more than it sounds, because every search happens inside an answer you're waiting on. Assistly doesn't search for things it already knows — a definition, an algorithm, arithmetic, code it can simply write. It doesn't search for anything already answerable from your screen or the conversation, it doesn't search the web for your own past meetings (that's the session search, which reads your history), and it doesn't search for you personally — the open web doesn't know you, and a search for a private individual returns a stranger.
It does search when you push back. "Are you sure?" or "I think that changed" is a cue to go and check rather than to restate the same answer more firmly, or to fold and agree with you when you were the one who was wrong.
And search results are treated as what they are: text from public web pages, quoted as source material. A page that happens to contain instructions is a stranger's writing, not you talking — Assistly reads it for facts and nothing else, and won't state a figure the results don't contain.
Neither one costs you time
Both of these could easily have made the assistant slower, and on a live call a slower answer is a worse answer no matter how well informed it is.
Memory is read from a cache that's refreshed when you change something, so a turn with memory costs nothing measurable against one without. The judgment about whether something you typed is worth remembering runs after your answer has finished — spending time while you're reading, never while you're waiting. And web search only runs when the model decides the question needs it, which is the point of teaching it when not to.
What this adds up to
Separately these are two features. Together they're a change in what you have to carry.
You stop re-explaining how you like to be answered. You stop re-stating what your product costs or what your role is. You stop mentally discounting every answer about the outside world by "well, how old is this model?" The assistant holds the standing stuff, checks the fresh stuff, and shows you both — the memories it used above the answer, the sources it read inside it.
Everything it remembers is one list you can read, edit, and empty. Nothing it remembers is a mystery.
Download Assistly and tell it something once.